Browse Topic: Driver behavior

Items (762)
To explore the impact of guiding and warning visual combination factors at the entrance sections of highway tunnels on drivers’ visual characteristics and driving behavior, this study recruited 16 drivers to conduct on-road vehicle experiments at the entrance sections of the Yunling Tunnel’s left bore (with visual combination factors) and right bore (without visual combination factors). Seven visual characteristics and driving behavior indicators, including pupil diameter and vehicle speed, were collected and statistically analyzed. Representative indicators such as pupil diameter, standard deviation of fixation point position, and vehicle speed were selected to establish a trend surface model of visual characteristics and driving behavior. The results indicate that when driving at the entrance section of the left bore, drivers’ pupil diameter and fixation duration were significantly lower than those at the entrance section of the right bore. With the increase in the sweeping view angle, there was a more dispersed distribution of fixation points. Additionally, there were significant differences in the acceleration and lateral deviation of the driving vehicle, with the range of variation narrowing by 52.5% and 35.7%, respectively. The trend surface model results show that under the influence of visual combination factors, the reduction in drivers’ vehicle speed was smaller, and the impact of pupil diameter and standard deviation of fixation point position on vehicle speed was less pronounced. Overall, under the influence of visual combination factors, drivers’ visual characteristics showed significant changes, with improved speed control and manipulation levels, leading to more stable vehicle operation.
Ma, YanpengHuang, HeHuang, YongYuan, Chen
The way we drive has a big effect on how much energy electric cars use, so making better driving habits can help make electric cars use less energy. By utilizing a set of real EV driving data, this paper classifies and analyzes EVs from the perspective of energy consumption, and establishes an intelligent scoring system for EV driving behavior based on a decision tree model. Experimental results show that this method is able to successfully distinguish different driving behaviours and the critical driving behavior factors, such as vehicle speed, accelerator pedal change rate, etc., and braking behavior are identified. Use intelligent scoring to give driver suggestions; this way, they can improve on their driving techniques and lower their energy consumption.
Liang, YongkaiZhang, HaoLiu, YuYu, Hanzhengnan
This study aims to analyze the impact of spatial and aspatial factors on the safety driving behavior of motorcycle couriers in East Jakarta within the context of the gig economy. Both factors are integrated to clarify how spatial conditions and individual characteristics jointly shape couriers’ safety driving behavior. The Partial Least Squares Structural Equation Modeling (PLS-SEM) method was employed to examine the relationship between spatial and aspatial factors on safety driving behavior. Data were collected through questionnaires from 253 motorcycle couriers operating in three subdistricts in East Jakarta, namely Cakung, Pasar Rebo, and Pulo Gadung. The results show that safety driving behavior is significantly influenced by aspatial factors, particularly socioeconomic characteristics and personality traits. In contrast, spatial factors such as road conditions and daily activity patterns do not directly influence safety driving behavior, but exert indirect effects through the couriers’ personality traits.
Wahyuddin, YasserSitorus, Paldibo AlfriramsonPutri, KharuniaMaharani, Garnierita
Rigorous validation of SAE Levels 3 and 4 autonomous systems increasingly relies on simulation. However, the simulation-reality gap remains a challenge for human-in-the-loop assessments. This study empirically quantifies the behavioral fidelity of the Car-Learning-to-Act (CARLA) simulator by recreating specific real-world traffic scenarios using the high-precision exiD drone dataset. Twenty-five participants performed a series of maneuvers, including lane changes and time-critical cut-ins. Their performance was analyzed using Dynamic Time Warping (DTW), driver profiling, and Time-to-Collision (TTC) metrics. The findings reveal a clear distinction between relative and absolute behavioral validity. In strategic decision-making tasks, the simulation demonstrated remarkably high temporal fidelity. DTW analysis explained 94% of the trajectory variance. Participants initiated lane changes with an average lag of -9 frames (0.36 s) compared to naturalistic references. These results indicate that, despite the absence of peripheral optical flow, the simulator successfully elicits temporally correlated decision-making patterns suitable for assessing strategic driver intent. However, physical execution in reactive scenarios revealed significant absolute discrepancies. Although the high Pearson correlation (r ≈ 0.89) in velocity profiles proves that drivers recognize and react to hazards with realistic timing, their physical inputs were exaggerated. Participants displayed digital, over-modulated braking responses and maintained a negative safety bias of -11.26 m, a deviation attributed to the lack of vestibular g-force feedback and geometric minification. Furthermore, distinct driver profiles emerged. Risk-oriented participants exhibited a gaming effect by neglecting safety margins. In conclusion, while CARLA is highly valid for testing the temporal logic of driver interactions, absolute dynamics require calibration functions, such as force-feedback (pedal) tuning and visual deceleration cues like camera shake, to compensate for sensory limitations before it can be used for safety-critical validation.
Rebling, PatrickAlphan, MetehanNenninger, Philipp
As automation advances and occupants transition from active drivers to passive passengers, understanding how automated driving behavior is evaluated becomes increasingly important. While longitudinal and lateral vehicle dynamics are known to influence perceived comfort and safety, it remains unclear to what extent motion–perception relationships remain stable across urban traffic contexts. This study compares two real-world investigations of automated driving: a left-turn maneuver at a signalized intersection on a test track and a roundabout maneuver with a shuttle in public traffic. Both datasets include high-resolution vehicle dynamics and structured subjective ratings. A consistent objectification approach was applied to examine the transferability of motion–perception relationships across contexts. However, differences in vehicle platform, automation level, trajectory characteristics, and study design limit direct comparability and require cautious interpretation. Despite partially overlapping ranges in selected peak-based dynamic parameters, such as longitudinal acceleration, subjective comfort and safety ratings were consistently higher in the roundabout scenario. Furthermore, strong associations were observed between motion parameters and subjective evaluations in the intersection context (adj. R2 up to 0.891), whereas objective parameters showed only limited explanatory power in the roundabout scenario (adj. R2 ≤ 0.06). The results indicate that motion–perception relationships derived within a specific context may not be directly transferable across different traffic scenarios. The findings highlight limitations of globally derived motion-based evaluation models and underline the importance of validating objectification approaches across diverse operational environments.
Panzer, AnnaStrenge, EmmaIatropoulos, JannesHenze, Roman
Driver monitoring systems are an important component of active safety systems, continuously evaluating the driver’s state and issuing real-time warnings. As defined by the SAE Levels of Automation, driving tasks are increasingly transferred from the driver to the vehicle from Level 0 to Level 2, however, the driver remains fully responsible for monitoring the driving environment. Current implementations, such as driver drowsiness and attention warning, assess driver alertness, while advanced driver distraction warning ensures that the driver maintains visual focus. Nevertheless, these systems do not identify the specific objects or regions the driver is observing. This limitation motivates the presented research question: can an in-car monitoring system be integrated with external environment perception sensors to infer the driver’s field of view (FoV)? This paper presents a system consisting of a driver-facing camera and a front-view camera. Facial features, including gaze direction, head pose, and iris offset are extracted using computer vision techniques. These features, together with cropped eye images, are used as inputs to a multi-modal network. Training labels were generated using a driving simulator study with 16 participants who sequentially fixated on visual targets displayed on a front screen. Experimental results show that the proposed system can predict driver visual attention and approximate FoV with a mean pixel error of 35.40 px, enabling identification of the regions of the road scene observed by the driver in real time. This work provides a foundation for explicitly modeling driver perception and its correspondence with vehicle perception systems.
Ji, DejieLausch, HendrykFlormann, MaximilianHenze, Roman
This article presents a data-driven pipeline for autonomous-vehicle (AV) safety testing. The pipeline integrates real-world traffic observations with model-guided scenario expansion and safety-metric evaluation to enable an end-to-end AV safety testing framework, demonstrated on a canonical highway scenario. The framework enhances test diversity, realism, and coverage by generating statistically informed variants of observed driving behaviors. Key parameters such as vehicle speed, trajectories, and headways are extracted from naturalistic data and used to train a probabilistic model of traffic dynamics. Scenario variants are sampled from this model and encoded as behavior trees (BTs) for modular, simulation-ready execution. Each scenario is simulated using a consistent AV control configuration, and safety metrics such as minimum safe distance violation, minimum safe distance factor, time to collision, and aggressive driving are applied to evaluate safety outcomes independently of system-specific tuning. A case study based on the highD dataset (110,000+ trajectories) demonstrates the framework’s ability to generate realistic and safety-relevant scenarios, providing an initial demonstration of pipeline feasibility and metric-based evaluation. This initial study is intentionally scoped to a single scenario class and a simplified parametric model to isolate and validate the end-to-end integration of the pipeline.
Elshenawy, MohamedAboudina, AyaAbdelmotaleb, AnharAmr, MariamEl-darieby, Mohamed
Large language models (LLMs) have shown remarkable capabilities for perceiving driving environments and making interpretable, logical decisions for autonomous driving. However, their potential for more comprehensive driving strategies, especially concerning energy efficiency, remains underexplored. Most existing studies primarily focus on driving safety, which may inadvertently increase energy consumption. To address this issue, this study explores the use of LLMs as high-level controllers to jointly optimize driving safety and energy efficiency. A textual prompt is designed for the LLM, incorporating few-shot examples that describe scenarios, states, and actions. The LLM processes the scenario and state prompts describing the surrounding traffic environment. It generates a high-level control signal, which is then translated into low-level vehicle motion commands in a high-fidelity traffic simulator with realistic physics, vehicle dynamics, road slopes, and network topology. Experiments in campus-scale digital twin car-following scenarios demonstrate that the proposed LLM-based framework achieves an average reduction of 4.16% in energy consumption compared to the reinforcement learning paradigm, while maintaining driving safety and providing interpretable high-level decision-making. This study highlights the potential of LLMs for longitudinal eco-driving applications under the evaluated simulation settings, extending previous LLM-based autonomous driving research that primarily focused on safety to also consider energy efficiency.
Wang, HaoyuLi, ZhenningWang, SiyingZhou, ZijingZhang, XiangYang, ZhifengOu, Shiqi (Shawn)Qi, Hao
Identifying driving heterogeneity is critical for enhancing the strategy learning capabilities of autonomous driving systems, as well as improving their safety and efficiency. This research proposes a novel driving heterogeneity identification framework. The framework consists of three core processes: action phase extraction, action relationship modeling, and behavior heterogeneity identification. First, a rule-based segmentation method is employed to systematically decode and interpret the inherent variations in human driving behavior. Subsequently, an action relationship modeling method is introduced to characterize the temporal relations between the acquired action phases. Finally, to mitigate the inaccurate identification caused by the sparse distribution of critical driving events in long-sequence data, a semantic encoding method is applied to remap the driving behavior space. Experimental results on the Lyft level-5 dataset validate the effectiveness of the proposed framework, which outperforms multiple traditional clustering algorithms. This demonstrates its significant potential to enhance behavior detection and learning in personalized advanced driver-assistance systems (ADAS) and advanced autonomous vehicle (AV) design.
Yin, HuiZhang, QinyaoLi, XiaojianMo, Hangjie
Letter from the Guest Editor
Tylko, Suzanne
NHTSA is conducting research to evaluate the current state-of-the-art technology for lane departure warning (LDW) and lane-keeping assistance (LKA) technology. NHTSA is undertaking research to understand the nature of real-world lane departures and recovery behaviors. While some information about lane departures can be learned from crash datasets, the purpose of this work was to mine simulator datasets for lane departures, analyze them in greater detail than is possible from crash reports or naturalistic studies, and link their characteristics to driver drowsiness. The objective of the study was to determine whether there are differences in lane departure characteristics as a function of driver drowsiness. This research used a novel approach by combining data from six different driving simulator studies on driver drowsiness. The dataset included a sample of 380 drivers. Study drives occurred during overnight hours after periods of sleep deprivation, with participants being awake for at least 16 h prior to driving. Study drives ranged in duration from relatively short 45-min to nearly 4 h. The datasets were reduced to characterize 5805 individual lane departures. Lane departures were delineated into three phases (pre-departure, departure, and recovery) and two transition points (onset and reentry) to capture driver behaviors under drowsiness. We hypothesized that lane departures would look different under different levels of drowsiness. Drives took place across a range of roadway environments that included interstate highways, rural highways, rural roads, and low-speed urban areas. Drowsiness was sampled at points before, during, and after the drive using self-ratings [Karolinska Sleepiness Scale (KSS) or Stanford Sleepiness Scale (SSS)] as well as the expert Observational Rating of Drowsiness (ORD). High levels of drowsiness were associated with a narrow speed range at highway speeds and the least amount of throttle input, while low levels of drowsiness had more steering activity, more throttle input, and a broader range of speeds. The results of this study will improve understanding of vehicle kinematics and driver behavior in drowsy lane departures using a safe methodology to help address crash dataset limitations.
Schwarz, ChrisGaspar, JohnShull, EmilyVenegas, Michael
Programs that teach older drivers how to confidently and competently use advanced vehicle technologies (AVTs) are limited. The MOVETech study evaluated a training program specifically designed to teach older drivers how to use these technologies. Participants (n = 119) were randomized to the intervention (training program) or control group (brochure). The intervention involved an in-person classroom education session on the use and benefits of AVTs, and an on-road driving session where participants drove along a pre-defined route in a dual-controlled vehicle with instruction on AVT use by a driving instructor. All participants completed in-person and telephone assessments at baseline and 3 months. Driving performance and on-road AVT competence assessments were the primary outcomes. Self-reported driving confidence, competence, and confidence in use of AVT, crashes, citations, and count of vehicle damage were the secondary outcomes. Program fidelity was also evaluated using a checklist. At 3 months, overall driving performance was high (96/100) and similar between groups. The intervention group, however, had slightly higher competence in AVT use (77 vs 73), but the between-group difference was not statistically significant (4.14, 95% CI −4.85 to 13.13). There were no differences in secondary outcomes. Program fidelity was high for all classroom sessions but varied for on-road sessions due to external and environmental factors, which impacted how AVT was demonstrated. The findings indicate AVT competence and confidence may be improved by combining classroom and on-road sessions, and importantly, that this type of program is feasible and very well-accepted among older drivers. Future work could target drivers with new vehicles who are unfamiliar with AVT to determine potential real-world benefits. This study provides evidence for vehicle manufacturers and policymakers to explore efficient ways of providing support to older drivers with AVT.
Nguyen, HelenRen, KerrieCoxon, KristyNeville, NickO’Donnell, JoanCheal, BethBrown, JulieKeay, Lisa
While an enlarged lead time from risk notifications to collisions is widely acknowledged to facilitate safe driving, it remains challenging to effectively notify drivers of invisible risks and non-apparent risks coming from uncertain behaviors on the part of road users. The current study examined whether verbal notifications are able to assist early awareness of predictive risks. We also attempted to identify human and environmental factors that could possibly improve the effectiveness of predictive risk information. Twenty-eight licensed drivers participated in a public road test conducted in two different urban areas on 3 days. They drove predefined courses on which potential risk locations were identified prior to the test, using a sport utility vehicle equipped with an automatic verbal notification system triggered based on the distance to the potential risk locations. After passing through the locations each time, the participants were instructed to verbally evaluate the shift in awareness provided by the notification and the usefulness of the assistance. After the driving test was completed, we acquired a subjective evaluation on annoyance acceptability and a self-report of participants’ road usage frequency at notified locations in daily life, as well as questionnaires on their driving style and workload sensitivity. We found that the effectiveness of verbal notifications increased by conveying uncertainty risks at visible locations and by using interrogative sentences or expressions of risk target perspective, although it decreased as a function of age. Our model showed strong performance in predicting positive ratings for the notifications, but this was not the case for negative ratings. We identified individual characteristics and the risk factor of uncertainty as important features in our model. In conclusion, the findings provide an important reference for understanding the early notification of predictive risk and constructing a numerical model for the implementation of assistance systems in vehicles and nomadic devices.
Maruyama, MasakiKoyama, KeiichiroEzaki, ToruSakamoto, JunichiSawada, YutaMatsuoka, Takahiro
Traffic collision reconstruction traditionally relies on human expertise and, when performed properly, can be incredibly accurate. However, attempting to perform pre-crash reconstruction, i.e., reconstructing the driver and vehicle behaviors that preceded the actual crash, poses significantly more challenges. This study develops a multi-agent artificial intelligence (AI) framework that reconstructs pre-crash scenarios and infers vehicle behaviors from fragmented collision data. We present a two-phase collaborative framework combining reconstruction and reasoning phases. The system processes 277 rear-end lead vehicle deceleration (LVD) collisions from the Crash Investigation Sampling System (CISS; 2017–2022), integrating textual crash reports, structured tabular data, and visual scene diagrams. Phase I generates natural language crash reconstructions from multimodal inputs. Phase II performs in-depth crash reasoning by combining these reconstructions with the temporal event data recorder (EDR). This enables precise identification of striking and struck vehicles while isolating the EDR records most relevant to the collision moment, thereby revealing crucial pre-crash driving behaviors. For validation, we applied it to all LVD cases, focusing on a subset of 39 complicated EDR cases where multiple EDR records per collision introduced possible ambiguity (e.g., due to missing or conflicting data). Ground truth was established via consensus between manual annotations (two independent researchers), with a separate large language model (LLM) used only to flag possible conflicts for re-checking. In the full end-to-end evaluation, the framework achieved 100% accuracy across all 4155 trials (277 cases × 5 runs × 3 models), with three reasoning models producing identical outputs, confirming that performance derives from the structured prompt design rather than model-specific characteristics. In contrast, research analysts without specialized reconstruction training achieved 92.31% accuracy on the same 39 complex cases. In separate ablation experiments on the 39 complicated EDR cases, where one randomly selected Phase I output from the full end-to-end evaluation was fixed as the unified input for Phase II and each model was tested with 10 independent runs, removing the structured reasoning anchors reduced case-level accuracy from 99.7% to 96.5%, with errors spreading from a single output type to multiple analytical dimensions. The system maintained robust performance even when processing incomplete data. This zero-shot evaluation, conducted without any domain-specific training or fine-tuning, demonstrates that the framework’s effectiveness stems from its multi-agent architecture and prompt engineering, offering a scalable approach for AI-assisted pre-crash analysis.
Xu, GeruiChen, BoyouGuo, HuizhongLeBlanc, DaveKusari, ArpanYarbasi, EfeAhmed, AnannaSun, ZhaonanBao, Shan
Drivers frequently encounter Type II dilemma zones at signalized intersections, where the decision to stop or proceed during the onset of a yellow indication can be ambiguous. Decision-making relies on drivers’ expectations of the yellow change interval duration and behavioral factors. While boundaries of these zones are well studied, less is known about how familiar drivers are with their local yellow indication laws, which vary from state to state, and whether their typical reactions to yellow indications align with the laws. Existing interventions like signal timing adjustments, improved vehicle detection, and advance warning signs reduce the number of drivers caught in dilemma zones but may not reach distracted drivers. In-vehicle alerts tailored to dilemma zone scenarios are a potential solution not yet implemented widely in North America. This study addresses how drivers may interpret these alerts. A web-based survey of 640 licensed drivers in Michigan and Washington (ages 18–85) assessed respondents’ knowledge of their state’s law, typical responses to yellow indications, interpretations of proposed in-vehicle alerts, and preferences for alert modality, frequency, and placement. These states were selected for their differing yellow indication laws—restrictive in Michigan, permissive in Washington. Nine alert visuals were tested, including pairs of implicit and explicit messages, and were inspired by or designed to address gaps in prior research. Respondents evaluated these alerts in response to hypothetical intersection scenarios that varied by the presence of other vehicles. Results revealed a prevalent misunderstanding of local yellow indication laws across both states. Statistical analyses showed significant differences in rankings among the nine alert visuals, and explicit messages showed higher rates of correct interpretation. Findings show overall driver support for dilemma zone alerts, but higher receptivity in drivers who more frequently use other ADAS features and lower receptivity in drivers within older, but not the oldest, age groups. Future research could explore whether these alerts promote safe behaviors aimed at crash avoidance.
Anderson, ErikaJashami, HishamAhmed, AnannaHurwitz, David
Vehicle maneuver data are essential for perception and planning in advanced driver-assistance systems (ADAS) and automated driving systems (ADS). While high-quality annotations improve machine-learning performance, existing maneuver datasets remain fragmented, labor-intensive to annotate, and inconsistent in semantic richness. Challenges persist in scalability, interpretability, and contextual labeling. This article establishes a structured framework for maneuver data analysis by combining a systematic review of existing resources with the development of a new multimodal dataset. First, we conduct a systematic review of publicly available datasets such as HDD, KITTI, BDD-X, D2CAV, Brain4Cars, DrivingDojo, and the Driving Behavior Database. We further evaluate the data modality and sensor configurations including event data recorders, onboard logging systems, and smartphone sensing. We then propose the Matt3r Data Collection System with modern metadata management, which integrates video, GPS, and IMU signals into temporally coherent clips. Next, we outline the limitations of traditional annotation approaches, which rely on manual labeling and rule-based methods. To address the limitations of traditional manual and semi-automated labeling, we propose a Vision–Language Model (VLM)–driven annotation pipeline. VLMs generate maneuver categories and causal explanations through prompt-based reasoning, with selected outputs refined through human-in-the-loop verification. Finally, we propose an annotation quality evaluation based on accuracy, inter-annotator agreement, credibility, consistency, and efficiency gain. In summary, this article bridges the gap between the environment perception requirements of existing ADAS and ADS systems and the developing capabilities of generative artificial intelligence. By providing a novel and scalable research approach for AI-driven maneuver data annotation and analysis, this article supports data engineering efforts for both research and practical applications aimed at enhancing vehicle safety.
Bai, LingYuan, ChongyuOsman, IslamLin, ZiruiMirab, GhazalSaheb, AmirParnian, NedaShapiro, EvgenyShehata, Mohamed S.Liu, Zheng
This study analyzed driver behavior in Turn-In-Path (TIP) scenarios using the Second Strategic Highway Research Program (SHRP2) naturalistic driving dataset. A total of 167 real-world incidents, including both crashes and near-crashes, were examined to evaluate human driver perception-response times (PRT) and avoidance behaviors when an intruding vehicle (the principal other vehicle, or POV) turns into the path of a straight-moving subject vehicle (SV). The combined analysis includes TIP events involving POVs turning from intersecting roads to either cross or merge into the SV’s lane and continues in the direction of the SV. Each event was reviewed to identify the driver behavior in an emergency response event, with measurements taken from video and telematics data. Response time was measured across two different starting points. Key variables included time to conflict, POV behavior, SV driver engagement in secondary tasks, and environmental factors such as lighting and roadway geometry. Across both datasets, shorter time to contact was consistently associated with quicker driver responses. Driver responses were slower when the POV entered from the right as well as when drivers were engaged in visual-manual secondary tasks. In contrast, driver age and gender were not found to significantly affect PRT. This combined study expands the understanding of real-world driver response behavior in TIP scenarios and provides an empirical foundation for refining crash avoidance systems and modeling human performance in traffic conflict situations.
Dinakar, SwaroopMuttart, JeffreyMaloney, TimothyAdhikari, Bikram
This study develops a personalized driver model for expressway merging, embedding individual driving characteristics into automated longitudinal and lateral control via Long Short-Term Memory (LSTM) networks. Uniform assistance (Advanced Driver Assist System, ADAS) can feel uncomfortable when it does not match a driver’s style; we therefore target the merge maneuver—a safety-critical task requiring anticipation and timing—and test whether merging-related context improves model fidelity. Driving data were collected in a high-fidelity motion-base simulator across two merging scenarios (13 licensed drivers in total). Inputs comprised ego speed, Headway distance and relative speed to the lead vehicle, and geometric context variables (distance to the end of the acceleration lane and to the hard/soft nose); outputs were longitudinal and, in the cross-scenario study, lateral accelerations. Models were trained per driver and evaluated by root mean square error (RMSE). Including merging context reduced longitudinal error in Experiment 1 (Gotemba IC) by about 30% on average relative to models without context, while errors remained below 0.5 m/s2. In Experiment 2 (Tokyo–Nagoya Expressway vs. Tokyo Metropolitan Expressway), longitudinal and lateral errors were low across both geometries; group-mean trends favored context but were non-significant, reflecting small sample size and inter-individual variability. Questionnaire-based evaluations in the simulator showed ratings close to real driving for discomfort, merge timing, and perceived safety; similarity and willingness to use were slightly higher in the urban expressway scenario, suggesting good user acceptance in constrained conditions. These findings indicate that incorporating merging context enables personalized control that better reflects individual driving behavior, while pointing to future work on generalization across geometries, speed ranges, and richer interaction semantics.
Shen, ShuncongHirose, Toshiya
As the adoption of electric vehicles continues to accelerate, the demand for their development and testing using chassis dynamometers has also increased significantly. Compared with internal combustion engine vehicles, chassis dynamometer testing for electric vehicles typically requires test durations several to several dozen times longer, resulting in substantially increased labor requirements. In addition, low-temperature testing is often required, further intensifying the workload associated with vehicle testing. To address these challenges, this study developed and evaluated a pedal robot designed to enable unmanned and automated testing. The pedal robot developed in this study weighs only 12 kg and can be installed within a few minutes. It is, to the authors’ knowledge, the world’s first pedal robot that mimics human driving behavior by using a single foot to operate both the accelerator and brake pedals. Unlike conventional driving robots, the actuators of the proposed system do not require direct mechanical attachment to the vehicle pedals, allowing for rapid installation. Furthermore, the robot is mounted on the driver-side floor, eliminating the need for attachment to the seat structure. The pedal robot features three degrees of freedom driven by three motors and employs artificial intelligence to recognize the shape and position of pedals across different vehicle models, thereby enabling automated test initiation without manual adjustment. The performance of the pedal robot was evaluated under UDDS, HWFET, and WLTC driving modes, and the results were analyzed in accordance with the SAE J2951 standard. Comparative evaluations demonstrated that the pedal robot achieved superior speed-tracking performance relative to that of an experienced human test driver. The developed pedal robot is currently being utilized for vehicle certification testing of electric and other vehicles at the Mobile Environment Research Center of the National Institute of Environmental Research in Korea. This paper presents a detailed analysis of the corresponding experimental results.
Lee, DaeyupKang, Ji MyeongJo, YechanChoi, SeongUnShin, JaesikKim, JongminKang, Keonwoo
Energy efficiency and range optimization remain critical challenges to the widespread adoption of battery electric vehicles (BEVs). As a result, there is a growing demand for intelligent driver assistance systems that can extend the operating range and reduce range anxiety. This paper presents an adaptive eco-feedback and driver rating system based on proximal policy optimization (PPO) reinforcement learning, designed to support drivers with the target to reduce energy consumption and maximize driving range. The system processes real-time driving data, such as velocity, acceleration and powertrain status. Map data of high quality is used to anticipate traffic events, including but not limited to speed limits, curves, gradients, preceding vehicles and traffic lights. This contextual awareness allows the system to continuously assess driving behavior and provide personalized, context-aware visual feedback alongside a dynamic driving behavior rating. A PPO agent learns optimal feedback strategies through continuous interaction and evaluates the impact of specific guidance actions, such as but not limited to “release accelerator pedal”, “brake” and “recuperate”, on immediate energy efficiency and long-term driver adaptation patterns. Feedback intensity and modality are dynamically tailored to individual driver profiles based on observed reaction patterns and feedback adherence. This approach encourages drivers to prioritize energy efficiency while aiming to minimize cognitive distraction and discomfort. The algorithm is implemented and validated within a driving simulation environment that replicates diverse and realistic conditions. Virtual driving tests conducted in various scenarios, such as congested urban areas, suburban routes, mountain roads and highways demonstrate that the proposed PPO-based eco-driving assistance system can reduce energy losses by about 28% compared to conventional driving behavior.
Stocker, ChristophHirz, MarioMartin, MichaelKreis, AlexanderStadler, Severin
Drivers often interact with partial automation (SAE Level 2) systems, initiating transfer of control (TOC) either by handing control over to the automation or by taking it back. Accurately predicting these interactions may inform the design of future automation systems that adapt proactively to the operating context, enhance comfort, and ultimately may improve safety. We present a context-aware framework that generates a unified driver–vehicle–environment representation by fusing data from in-cabin video of the driver and of the forward roadway with vehicle kinematics, driver glance, and hands-on-wheel behaviors. This representation was encoded in a hierarchical Graph Neural Network that classified driver-initiated TOCs to: (i) Manual-to-automation and (ii) Automation-to-manual transitions and predicted time-to-TOC. Shapley-based explainable AI was used to quantify how the importance of behavioral, contextual, and kinematic cues evolved in the seconds preceding a TOC. Analysis of a naturalistic dataset of 1,565 driver-initiated TOCs from 16 experienced drivers revealed distinct patterns. Manual-to-automation transitions were preceded by lane count increases, acceleration, and spikes in glances to the instrument-cluster. In contrast, Automation-to-manual transitions were associated with lane count reductions, higher surrounding-vehicle density, deceleration, reduction in secondary-task engagement, and higher steering wheel control. Together, these patterns highlight key cues for predicting the TOC type and time-to-TOC. Using environment-only features, the classifier achieved 78% accuracy; adding vehicle kinematics increased accuracy to 84%, and incorporating driver behavior features further improved prediction to 90%. Across prediction horizons, the Manual-to-automation TOC was consistently predicted more accurately than the automation-to-manual TOC. Shapley analyses underscore that driver behavior provided the strongest cues for predicting TOCs, highlighting the value of fusing driving context with information obtained from monitoring the driver behavior to anticipate the type of driver-automation interaction and its timing.
Zhao, ZhouqiaoGershon, Pnina
Autonomous vehicle navigation requires accurate prediction of driving path curvature to ensure smooth and safe trajectory planning. This paper presents a novel approach to curvature prediction using deep neural networks trained on GPS-derived ground truth data, rather than model predictions, providing a more accurate training signal that reflects actual vehicle motion. We develop a multi-modal neural network architecture with temporal GRU encoders that processes vision features, driver intent signals, historical curvature, and vehicle state parameters to predict curvature. A key innovation is the use of GPS-based actual curvature measurements computed from vehicle motion data (κ = ωz/v) as training supervision, enabling the model to learn from real-world driving patterns. The model is trained on 5,322 samples from real-world driving data collected on The University of Oklahoma’s Norman Campus using a Comma 3X device and a 2025 Nissan Leaf electric vehicle. Experimental results demonstrate high steering curvature prediction accuracy with a Pearson correlation coefficient of 0.805, Mean Absolute Error of 0.027654, and Root Mean Squared Error of 0.034402 on the validation set. The model achieves stable convergence within 10 epochs and maintains consistent performance across diverse driving scenarios, from straight highway segments to complex turning maneuvers. This work contributes to autonomous driving technology by demonstrating the effectiveness of GPS-supervised learning for curvature prediction, successfully deployed in OpenPilot’s production system with real-time inference at 5 Hz.
Hajnorouzali, YasamanWang, HanchenLi, TaozheBurch, CollinLee, VictoriaTan, LinArjmandzadeh, ZibaXu, Bin
Regenerative braking has a strong influence on the energy efficiency and drivability of battery-electric vehicles. This study establishes an empirical baseline analysis under controlled conditions of the regenerative braking behavior of the 2020 Tesla Model 3 to support the interpretation of on-road performance and serve as a reference for subsequent testing and analysis. The tests were performed on a four-wheel-drive chassis dynamometer at Argonne National Laboratory, combining Multi Cycle Testing (MCT) to simulate real world driving patterns (city, highway) with coast-down tests to isolate periods where the motor is operating in regen mode and compare the behavior across different parameters. Vehicle data was collected from the vehicle using taps in the Controller Area Network (CAN) bus as well as a high-resolution power analyzer. The vehicle displayed the highest efficiency during simulated city driving conditions (3.62 miles/kWh followed by highway (3.40 miles/kWh) and aggressive (2.53 miles/kWh) conditions, though aggressive driving showed the highest energy recovery. Regenerative energy recovery was most efficient in the 10 – 30 mph range, with the rear motor regenerating all the energy while the front motor used a small amount of power. Standard regen mode achieved 57% greater deceleration during coast down compared to Low Regen mode and showed a much lower variability during different simulated uphill and downhill conditions. Standard mode collected more energy than Low mode in all cases apart from simulated downhill tests where Low mode performed better. These results provide an overview of the Tesla Model 3 regenerative braking behavior and delineate operating regimes that maximize efficiency and quantify trade-offs between deceleration stability and energy recovery across driver-selectable modes. The results provide a rigorous, reproducible baseline and measurement protocol that can enable cross-vehicle benchmarking, validate vehicle/software-in-the-loop models, and inform future controller calibration and the design of on-road and track experiments
Pierce, Benjamin BranchDi Russo, MiriamDas, DebashisZhan, LuStutenberg, Kevin
Electric vehicles (EVs) are central to sustainable transport, yet battery service life remains a limiting factor for cost and adoption. Distinct from traditional laboratory-based simulations that often fail to capture the complexity of field conditions, this study investigates how EV user behavior—including driving style and charging demands—influences capacity using large-scale, real-world operational data from daily EV usage. A data-driven framework is developed to quantify driving and charging behaviors through multidimensional feature extraction at the vehicle level and estimate battery State-of-Health (SOH) trajectories, enabling direct linkage between individual behavior patterns and degradation outcomes. Results reveal substantial heterogeneity in aging rates explicitly driven by diverse user behaviors: under identical urban conditions, vehicles with a radical driving style exhibit approximately 81% faster SOH decline per 20,000 km than those with a moderate style; regarding charging intensity, in controlled comparison scenarios, increasing fast-charge counts from the baseline interval of 90–120 to the elevated interval of 150–180 is associated with a ~2.1% reduction in median SOH when holding other factors constant; and similarly, increasing deep charge–discharge events from 160–180 to 220–240 corresponds to an additional ~2.0–2.3% cumulative SOH loss. These findings quantify the behavioral determinants of capacity fade in the field and demonstrate that aggressive driving, frequent fast charging, and deep discharge habits materially accelerate battery degradation. The framework provides actionable evidence for adaptive charging guidance and personalized driving strategies, extending battery longevity and enhancing the sustainability of electric mobility systems.
Liu, TianyiJing, HaoZhu, JiankuanChen, YongjianOu, ShiqiQian, Xiaodong
In recent years, the tightening of vehicle emission regulations has led to a decreasing trend in regulated pollutants such as NOₓ and CO. However, the emission of ammonia (NH₃), which is unintentionally generated during the purification process in three-way catalyst of gasoline vehicles, has become a growing concern. NH₃ emissions from vehicles can serve as a precursor to PM2.5 and have been reported to cause local roadside pollution. Therefore, there is a growing need for on-road testing to identify conditions under which NH₃ is likely to be emitted. Furthermore, since engine control strategies vary among vehicle types, it is desirable to consider differences in emission behavior across different models. In this study, on-road NH₃ emissions were measured for multiple vehicle models with different powertrains, and the effects of engine behaviors and engine operating duration across vehicles on NH₃ emissions were investigated. To analyze differences in NH₃ emission behavior among vehicle types, conventional gasoline vehicles and series-type hybrid vehicles were employed. Additionally, vehicle control parameters were obtained via an OBD (On-Board Diagnostics) interface unit and utilized for analysis. The analysis revealed that, for the conventional gasoline vehicles, aggressive accelerator pedal control induced rapid fluctuations in engine speed, which in turn led to NH₃ emissions. In contrast, for the series-type hybrid vehicles, NH₃ emissions were primarily observed when the engine started under specific conditions, whereas differences in driver behavior had only a minor direct impact on NH₃ emissions. In addition, longer engine operating durations resulted in higher emission levels. A common characteristic observed across both vehicle types was that NH₃ emissions were elevated during periods corresponding to CO emissions, which serve as precursors to NH₃ formation.
Ashizawa, KeigoFukunaga, ChisatoGao, TianyiSato, Susumu
Head-on emergency events present unique challenges for evaluating both human and automated-vehicle (AV) performance because they do not conform to a direct stimulus–response sequence. Instead, driver behavior in these scenarios follows a stimulus–wait–response pattern governed by time-to-conflict (TTC), uncertainty, and environmental affordances. Prior research has often failed to distinguish between conflict types, resulting in generalized reaction-time assumptions that do not account for contextual uncertainty. This study integrates simulator and naturalistic driving data from a four-part research program to establish objective benchmarks for driver responses in head-on encounters. When an encroaching vehicle crossed the centerline 2.5 s before impact, drivers initiated braking with a weighted average of approximately 1.0 s before impact. When the encroaching vehicle crossed or was first observed at approximately 3.5 s before impact, braking typically began with a weighted average of 1.3 s before impact, consistent with a deliberate waiting period rather than an immediate reaction. Across conditions, response variability increased with TTC, with the standard deviation scaling at 0.50 times the mean. In events where the encroaching driver corrected back to the proper lane, delayed responses were associated with successful avoidance. Steering behavior was influenced by roadside affordances, drivers steered right when no right-side obstacle was present but rarely steered right when any obstacle existed and drivers were likely to steer left when right-side obstacles were present. These findings reinforce the wait-and-see principle and provide empirically grounded benchmarks for evaluating human and AV responses in head-on emergency scenarios.
Muttart, JeffreyDinakar, SwaroopMaloney, TimothyAdikhari, BikramGernhard-Macha, Suntasty
Despite remarkable advances in vehicle technology - enhancing comfort, safety, and automation – productivity of transportation over the road continues to decline. Stop-and-go driving remains one of the most persistent inefficiencies in modern mobility systems, leading to greater travel delays, energy waste, emissions, and accident risk. As vehicle volumes rise, these effects compound into systemic challenges, including driver frustration, unstable flow dynamics, and elevated greenhouse gas (GHG) emissions. To address these issues, an extensive data-driven evaluation was performed characterizing the underlying causes of traffic instability and uncovering hidden behavioral parameters influencing traffic flow. This research led to the identification of a previously unrecognized metric - the Driver Comfort Index (DCI) - which quantifies an inter-vehicle spacing behavior that reflects intrinsic human driving behavior. Building on this discovery, mixed traffic is explored to identify its phenomena, where human-driven and machine-controlled vehicles coexist and share the road. It appears that adaptive cruise control (ACC) and connected autonomous vehicles (CAV) are controlled by a non-intrinsic parameter so that traffic mix suffers from a mismatch of vehicle dynamics. This mismatch is explored, and it is proposed to harmonize traffic dynamics by adopting the natural DCI parameter as the single control mechanism. Analytical studies demonstrate that DCI-based traffic flow orchestration, applied integrally to human- and machine-controlled vehicles, enhances traffic flow stability, mitigates stop-and-go oscillations, and significantly improves network efficiency, safety, and environmental performance.
Schlueter, Georg J.
Avoiding and mitigating any potential collision is dependent on (1) road user ability to avoid entering into a conflict (conflict avoidance effect) and (2) road user response should a conflict be entered (collision avoidance effect). This study examined the collision avoidance effect of the Waymo Driver, a currently deployed SAE level 4 automated driving system (ADS), using a human behavior reference model, designed to be representative of a human driver that is non-impaired, with eyes on the conflict (NIEON). Reliable performance benchmarking methodologies for assessing ADS performance are an essential component of determining system readiness. This consistently performing, always-attentive driver does not exist in the human population. Counterfactual simulations were run on responder collision scenarios based on reconstructions from a 10-year period of human fatal crashes from the Operational Design Domain of the Waymo ADS in Chandler, Arizona. Of 16 simulated conflicts entered, 12 (75%) were prevented by the Waymo Driver, and 10 (62.5%) were prevented by the NIEON model. The NIEON Model mitigated an additional 5 collisions and did not mitigate 1 collision. In these 16 conflicts entered, 93% of serious injury risk was reduced by the Waymo Driver, whereas 84% of serious injury risk was reduced by the NIEON model. Further, in a case-by-case evaluation, the Waymo Driver’s collision avoidance led to reduced serious injury risk when compared to the NIEON model in every simulated event. The results of this paper demonstrate that a reference model like NIEON can be used to benchmark ADS responder performance in response to high-risk initiating behaviors performed by the current driving population.
Scanlon, John M.Kusano, Kristofer D.Engstrom, JohanVictor, Trent
A Detroit-based startup says its device can analyze brain activity to help figure out whether a driver is impaired. The impaired driver-detection business has been heating up since even before NHTSA announced in 2024 that it was working what would eventually be a mandate that vehicles be able to detect impaired drivers and mitigate the danger they represent to the motoring public.
Clonts, Chris
Driving behavior is a significant factor influencing vehicle emissions, and it must be carefully considered when modeling emissions for real road transportation vehicles. This study aims to contribute to this field by improving the intelligence and accuracy of distinguishing driving behavior volatility through the use of clustering algorithm. The research begins by processing raw emissions data collected from light-duty gasoline vehicle during real-driving emissions (RDE) test, which are used as input features for the clustering algorithm. Subsequently, a driving behavior classification method based on the gaussian mixture model (GMM) clustering algorithm is proposed. The results show that aggressive driving has a significantly higher CO2 emission rate compared to normal and calm driving. Specifically, the average CO2 emission rate for aggressive driving is 5.61 g/s, which is substantially higher than that of calm driving (2.40 g/s) and normal driving (2.91 g/s). Following this, the study employs Pearson correlation coefficients and an intelligent machine learning modeling approach, incorporating driving behavior–related parameters as part of the input features for the virtual CO2 emission prediction model for light-duty gasoline vehicles. This further highlights the effectiveness of using a clustering algorithm for driving behavior classification, and it demonstrates the significant impact of driving behavior–related parameters on CO2 emission rates, providing an accurate emission model for virtual calibration of vehicles.
Yu, HaoMa, YiTan, JianxunWang, JingZhang, HonghaoHu, WeiChen, HaoYu, Wenbin
Electric vehicles (EVs) are the cornerstone of sustainable transportation, but their performance and component longevity are heavily influenced by driving behaviors. This study proposes a comprehensive analytical framework to assess how different driving styles affect the operational health of key EV components such as the battery pack, motor, and DC-DC converter. Various driving styles such as aggressive, moderate, and economical are discriminated against using dynamic vehicle operation signatures including acceleration and braking intensity, turning profiles, and load variations. These behavioral patterns are reflected in the electrical responses, namely current and voltage waveforms across power electronic systems. By analyzing these electrical signatures, a range of KPIs can be estimated for each component, offering insights into their operational stress and degradation trends. Experimental analysis using real-time EV datasets validates the framework’s ability to predict and correlate driving patterns with component degradation trends. By evaluating these impacts, the study bridges the gap between driving behaviors and their consequences on EV performance in real-time operation. The research culminates in generating prescriptive and descriptive analytics, offering actionable insights and tailored recommendations for driving behavior improvement. Real-time suggestions empower drivers to adopt safer, more efficient styles, while actionable strategies provide a roadmap for EV manufacturers and policymakers to promote sustainable and efficient transportation systems. This study underscores the critical interplay between driving behavior and EV component health, paving the way for smarter, data-driven mobility solutions
Deole, KaushikKumar, PankajHivarkar, Umesh
In recent times, a standard driving cycle is an excellent way to measure the electric range of EVs. This process is standardized and repeatable; however, it has some drawbacks, such as low active functions being tested in a controlled environment. This sometimes causes huge variations in the range between driving cycles and actual on-road tests. This problem of variation can be solved by on-road testing and testing a vehicle for customer-based velocity cycles. On-road measurement may be high on active functions while testing, which may give an exact idea of real-world consumption, but the repeatability of these test procedures is low due to excessive randomness. The repeatability of these cycles is low due to external factors acting on the vehicle during on-road testing, such as ambient temperature, driver behavior, traffic, terrain, altitude, and load conditions. No two measurements can have the same consumption, even if they are done on the same road with the same vehicle, due to the influence of the above-mentioned external factors. The current paper will portray a machine learning-based methodology to parameterize the external factors affecting e-motor consumption. By parameterizing these factors, on-road test results are normalized and further used for comparative studies. The paper also takes us through the process of data collection for this study, the parameterization process of external factors using ML models, for different driving scenarios and ambient temperature ranges. The ML models are developed in a MATLAB environment and can be reproduced in any other tool. Merits and demerits of each ML model are discussed along with ways and means to mitigate each external factor, which will make the testing procedure more robust and reliable. Thus, it helps in making automobiles more energy efficient.
Kelkar, KshitijKanakannavar, Rohit
Identifying the type of drive cycle is crucial for analyzing customer usage, optimizing vehicle performance and emission control. Methods that rely on geographical location for drive cycle identification are limited by varying driving conditions at the same location (e.g. heavy traffic during peak hours vs. free-flowing traffic at night). This paper proposes a methodology to identify the type of drive cycle (city, interurban, highway or hybrid) using drive characteristics derived from vehicle data rather than geographical location. Real-world vehicle data from testing trucks is taken, whose drive profiles are already known. Initially, multiple characteristic features of the drive cycle are identified from literature surveys and domain experience. These features, which can be extracted from basic signal data, include gear shifts, time spent in different driving modes (acceleration, cruise, standstill), velocity distributions, and an 'aggressiveness factor' representing overall driving style. Using ML based feature selection techniques, the most important features are selected for this cause. With these finalized parameters, a data-driven classification model is developed. This model is trained, validated, and tested using the identified real-world vehicle data. It classifies drive cycles into four major types: city, interurban, highway, and hybrid with a high degree of accuracy. This classification enables accurate identification of drive cycles, addressing the limitations of location-based methods. The developed model is employed to determine the type of drive cycle driven by customers, thereby aiding in the analysis of the influence of drive cycles on vehicle performance and emissions.
Reddy, Mallangi PrashanthGorain, RajuGanguly, Gourav
In India, Currently Continuous FULL MIDC (Modified Indian Driving Cycle) is used to declare the Range & Energy consumption of BEV (Battery Electric Vehicle). AISC (Automotive Industry Standards Committee) is looking to implement Worldwide Harmonized Light-Duty Test Procedure (WLTP) in India. AISC released AIS 175 for WLTP implementation from Apr 2027. The objective of WLTP is to standardize the test procedure globally for evaluating Emission/FE/Range of Light Duty Vehicles. But the effect of AIS 175 regulation on Battery Electric Vehicles Range Declaration is very less. The Range is almost same as Full MIDC declared Range. The On-road Range BEV is always lesser than the Declared Range of vehicles because of ambient conditions. Usually, the Full MIDC declared Range will be 20% ~26% higher than actual On Road Range. The Range of BEV as per India WLTP 3-Phase was observed 18% ~ 24% higher than actual On-road range of vehicles. There is only 2% difference observed between Full MIDC Range test & India WLTP 3-Phase PER test results. The Energy Consumption of WLTP 3-Phase test is 7% better than MIDC Energy Consumption test (AIS 039). To develop and declare the range of BEV’s near to on-road range better WLTP test cycle is required. So, we created New India WLTP 4-Phase test cycle with 4Phases: Low, Medium, High1, High2 (reference from Europe WLTP, instead of Extra-High phase, High phase repeated, just like NEDC to MIDC). The Range observed in India WLTP 4-Phase is only 10% ~ 16% higher than on-road Range of vehicle. The difference between India WLTP 3-Phase vs India WLTP 4-Phase for Range and Energy consumption was 6% Implementing India WLTP 4-Phase test cycle will provide very realistic range estimation results for Indian driving conditions compare to India WLTP 3-Phase test cycle.
Shiva Kumar, MucharlaTentu, Kavya
In-vehicle communication among different vehicle electronic controller units (ECU) to run several applications (I.e. to propel the vehicle or In-vehicle Infotainment), CAN (Controller Area Network) is most frequently used. Given the proprietary nature and lack of standardization in CAN configurations, which are often not disclosed by manufacturers, the process of CAN reverse engineering becomes highly complex and cumbersome. Additionally, the scarcity of publicly accessible data on electric vehicles, coupled with the rapid technological advancements in this domain, has resulted in the absence of a standardized and automated methodology for reverse engineering the CAN. This process is further complicated by the diverse CAN configurations implemented by various Original Equipment Manufacturers (OEMs). This paper presents a manual approach to reverse engineer the series CAN configuration of an electric vehicle, considering no vehicle information is available to testing engineers. To execute reverse engineering, the CAN data log is to be taken from the OBD-II port by systematically identifying and mapping the CAN with various ECUs interfaced with that CAN line. Driver actions and continuous data logged from the OBD-II port are cross-referenced with CAN data to determine the byte order of signals and message frames. The signals derived from one vehicle use scenario (driving) are then validated against another scenario (charging) to ensure consistency and accuracy.
Kumar, RohitSahu, HemantPenta, AmarBhatt, Purvish
In the evolving landscape of the automotive industry, this study presents an innovative approach to developing digital twins for driver profiles, establishing a standardized and scalable procedure for collecting and analyzing driving data on a global scale. The proposed methodology centers on the development of a robust cloud infrastructure, including Data Lake and associated services, designed for efficient storage and processing of large volumes of data from multiple markets and vehicle types. The research introduces an adaptable procedure for data collection campaigns, applicable to diverse global markets and encompassing a wide range of vehicles, from internal combustion engines to electric and hybrid models. A key feature of this approach is the establishment of advanced data decoding protocols, enabling precise interpretation of CAN network information from vehicles of different manufacturers and models, even when the CAN structure is not previously known. The study defines standardized parameters for data recording, ensuring comparability across different markets and vehicle types, while developing adaptive analysis methodologies to identify specific driving patterns based on vehicle segment, propulsion technology, and demographic characteristics. This comprehensive approach is underpinned by a framework that ensures compliance with data protection regulations globally, facilitating ethical and legal data management across different jurisdictions. The anticipated outcomes include the creation of a highly flexible data-as-a-service platform capable of integrating and analyzing driver data worldwide, and the establishment of a standardized procedure for characterizing driver profiles. The development of advanced vehicular data decoding capabilities allows for the inclusion of a wide variety of brands and models, enabling truly global insights into driver behavior. This study lays the groundwork for a global understanding of driver behavior, providing automotive manufacturers with a powerful tool to adapt their designs to the needs of users worldwide, accelerating innovation in vehicle design and improving the safety of future vehicles.
Arturo, RubioMarín Saltó, AnnaDiaz, FranciscoOlivencia, Sergio
Bilateral Cruise Control (BCC) is a new concept that has been shown to reduce traffic congestion and enhance fuel/energy efficiency compared to Adaptive Cruise Control (ACC). BCC considers both lead and trailing vehicles to determine the ego vehicle’s acceleration, effectively damping any disturbance down the vehicle string and reducing possibilities for congestion. Despite the advantages demonstrated with BCC, one major limitation is its non-intuitive behavior, which stems from the fact that the BCC reacts not just to the lead vehicle but also to the trailing vehicle’s movement. This paper identifies key issues with BCC control and proposes solutions that retain the benefits of BCC while maintaining intuitive behavior. Specifically, a novel switching strategy is proposed to switch between ACC and BCC control modes by critically analyzing the driving conditions. The proposed system ensures acceptable driving behavior with predictable braking and acceleration, resulting in an intuitive and smooth traffic flow. Through seamless integration of ACC and BCC, the system can prevent traffic congestion problems while closely aligning with human driving expectations.
A, AryaA, AishwaryaD, Vishal MitaranM, Senthil VelKumar, Vimal
With introduction of Corporate Average Fuel Efficiency norms (hereafter referred as CAFÉ norms) in India, the manufacturers of all M1 Category vehicles (not exceeding 3,500kg GVW) must ensure that they comply with Annual Corporate average CO2 target as defined in regulation. Moreover, this target will become stricter at various stages in the coming years. Hence CO2 emissions are becoming one of the major focus parameters during vehicle development. There are several factors that can impact CO2 emissions during measurement in laboratory-based test cycles such as MIDC or WLTC. One such major factor is driving variations. Although speed and time tolerances are provided during the test (as part of AIS 137/AIS 175) to limit the variation, even within these tolerances, drive-related effects make significant contribution to test results variability. Monitoring and control of such variations is important to understand the true fuel economy potential of the vehicle. Drive Trace indices are standardized metrics that can be used to evaluate the driving variations. The aim of this study is to understand the different driving behaviors on drive indices and consequently on CO2. Drive indices such as Energy Rating (ER), Distance Rating (DR), Energy Economy Rating (EER), IWR (Inertial Work Rating), RMSSE (Root Mean Square Speed Error) defined in SAE J2951 document have been referred for this study. Multiple MIDC & WLTC emission test data have been used for evaluation of driving behavior. An attempt has been made to establish a correlation between the drive trace indices and CO2 (and fuel economy) in MIDC by using mathematical techniques similar to study done by JRC for WLTC.
ER, ShivramRawat, VijaypalKhandelwal, VineetKumar, ArunMalhotra, Jitendra
In automotive engineering, understanding driving behavior is crucial for decision on specifications of future system designs. This study introduces an innovative approach to modeling driving behavior using Graph Attention Networks (GATs). By leveraging spatial relationships encoded in H3 indices, a graph-based model constructed, which captures dependencies between various vehicle operational parameters and their operational regions using H3 indices. The model utilizes CAN signal features such as speed, fuel efficiency, engine temperature, and categorical identifiers of vehicle type and sub-type. Additionally, regional indices are incorporated to enrich the contextual information. The GAT model processes these heterogeneous features, learning to identify patterns indicative of driving behavior. This approach offers several significant advantages. Firstly, it enhances the accuracy of driving behavior modeling by effectively capturing the complex spatial and operational dependencies inherent in vehicle data. The use of GATs allows for the dynamic weighting of different features, ensuring that the most relevant information is prioritized in the analysis. Secondly, the integration of regional indices provides a deeper contextual understanding, enabling the model to discern region-specific driving patterns that might otherwise be overlooked. Furthermore, this method facilitates the identification of abnormal behavioral trends, offering valuable insights for design engineers. By understanding region-based driving behavior, engineers can modify vehicle systems to better meet the needs of specific areas, leading to improved performance and user satisfaction. The combination of graph-based methods with attention mechanisms represents a significant advancement in vehicle performance monitoring, paving the way for a more comprehensive understanding of driving behavior across different regions.
Salunke, Omkar
With the growing adoption of Advanced Driver Assistance Systems (ADAS) in the Indian automotive landscape, the need for effective Driver Monitoring Systems (DMS) has become increasingly critical. This paper presents the design, development, and validation of a Driver Distraction and Attention Warning System (DDAWS) tailored to Indian driving conditions. The proposed system integrates two key modules: Driver Attention Monitoring and Drowsiness Detection, using a high-resolution driver-facing camera to analyse head pose, facial landmarks, and behavioural cues. The drowsiness module incorporates metrics such as PERCLOS and Eye Aspect Ratio (EAR), evaluated against the Karolinska Sleepiness Scale (KSS). Recognizing the limitations of self-assessed scales like KSS in dynamic driving environments, the study compares algorithmgenerated KSS values with self-reported scores to assess model accuracy. Additionally, the framework aligns with automotive safety standards such as AIS184,EU 2021/1341, EU 2023/2590, and EURO-NCAP. A multi-level redundancy architecture is introduced to improve prediction robustness by fusing outputs from both attention and drowsiness subsystems. The result is a scalable, regulation-compliant, and reliable DDAWS framework, optimized for real-world deployment in Indian vehicles.
Verma, HarshalKale, Jyoti GaneshKarle, Ujjwala
In the Indian context, introduction of ADAS can play a positive role in improving road safety by assisting the driver and preventing unsafe driver behaviour. Technologies like Automated Emergency Braking (AEB), Lane Keep System, Adaptive Cruise Control, Driver Drowsiness Detection, Driver Alcohol detection etc., if deployed safely and used in a safe manner can help prevent many of the current road deaths in India. Safe deployment and safe use of such ADAS technologies require the systems to operate without failure within their operational design domains (ODD) and not surprise the drivers with sudden or unpredictable failures, to help develop their trust in the technology. As a result, identifying test scenarios remain a key step in the development of Advanced Driver Assistance Systems (ADAS). This remains a challenge due to the large test space especially for the Indian context due to the unpredictable traffic behaviour and occasional road infrastructure. In this paper, we introduce a novel open-access crowd-sourcing public platform, Safety Pool™ Studio, to enable crowdsourcing of traffic scenarios in the Indian context. Safety Pool™ Studio platform enables any member of the public or the road traffic ecosystem (e.g. traffic police, local authorities, academia etc.) to create a traffic scenario using a graphical interface, like a LEGO making exercise. This would enable the users to share their real-life experiences of traffic scenarios in a simple, accessible and inclusive manner and contribute to a global pool of traffic scenarios in the Indian context. Safety Pool™ Studio provides multi-language support for India’s regional languages like Hindi, Bengali, Tamil, Marathi, Punjabi, Kannada, Telugu, Gujrati among others. Safety Pool™ Studio has been developed in a way the graphical scenarios can automatically be converted into programmatic description of scenarios for traditional simulation-based testing of ADAS.
Serry, HamidDodoiu, TudorAlakkad, FadiZhang, XizheKhastgir, SiddarthaJennings, Paul
In India, fuel economy is one of the most critical factors influencing a customer's decision to own a passenger car. Beyond consumer preference, fuel consumption also plays a significant role in the nation's energy security. In line with this, the government promotes fuel-efficient vehicles and technologies through various regulations, policies, and mandates. Vehicle manufacturers, in response, focus on designing vehicles that align with both customer expectations and regulatory requirements. Fuel economy certification is typically based on standardized laboratory tests that simulate controlled environmental conditions, driving cycle (MIDC), vehicle load, and operation of electrical and electronic systems. However, actual on-road driving conditions by end user vary significantly due to factors such as traffic conditions, ambient temperature, air conditioning use, driving behavior and variable loading of the vehicle. With implementation of Bharat Stage VI, Real Driving Emission (RDE) became mandatory from April 2023 to meet the requirements of conformity factors (CF) for NOX and PN emission. RDE regulation scope doesn’t include measurement or compliance for fuel economy during real driving condition. For the purpose of this study, laboratory and real driving emissions (RDE) testing were carried out in accordance with AIS 137 Part 3. For systematic comparison, fuel economy was calculated after modifying Carbon Balance equation in line to CAFÉ regulation S.O. 1072 (E) Dated 23rd April 2015. This study presents a comparative analysis of fuel economy results obtained from the testing different vehicles operating on different fuels like Gasoline, Diesel and Bi-fuel (Compressed Natural Gas (CNG) + Gasoline). The paper concludes with finding of study as impact of real-world driving conditions, particularly of ambient temperature and real driving on fuel efficiency of passenger cars.
Singh, Abhay PratapBathina, Revanth KumarTijare, Shantanu
Passenger vehicle users often manoeuvre their cars in diverse and unpredictable driving patterns. The vast and varied terrain of the Indian subcontinent further complicates this scenario, introducing unique challenges due to differences in driving expertise, vehicle usage, and environmental conditions. A specific challenge addressed in this paper arises during different engine temperatures and transient driving conditions—a critical phase for engine calibration to ensure optimal drivability and emissions performance. With current calibration practices, a backfire like abnormal engine noise was observed during certain transient driving patterns. This paper presents a novel calibration methodology designed to eliminate such abnormal noise. The proposed approach involves a step-by-step transient calibration refinement, making the calibration process more robust and adaptable to any driving behaviour. The paper outlines the specific challenges encountered and details the multi-level calibration and validation strategy used to resolve the issue, thereby enhancing overall vehicle drivability performance and customer satisfaction.
Suna, BhagyashreeTyagarajan, SethuramalingamPise, ChetanAishwarya, Amritansh
In automotive safety systems, Time to Collide (TTC) is traditionally used to trigger warnings in auto-emergency braking systems. However, TTC can lead to premature or inaccurate warnings as it is calculated based on the relative speed and distance between the ego and an obstacle. TTC does not consider the vehicle’s braking dynamics, such as brake prefill lag which varies across different vehicles, maximum deceleration, and the effectiveness of braking systems and assumes constant speed which may not always be realistic. We propose Time to Brake (TTB) as a more effective parameter for driver warnings. TTB directly relates to the action a driver needs to take—braking. It provides a clear indication of when braking should begin to avoid a collision, whereas TTC only tells us about the possibility of a collision. To calculate TTB we utilize the brake profile, which incorporates both deceleration and system jerk for improved accuracy. The proposed warning time is the sum of variable brake prefill lag, average driver reaction time, and TTB. TTB is calculated for two distinct scenarios due to differing constraints using Newtonian equations of motion. The comoving and oncoming scenario involve both ego and object colliding at the same location, but in the former, the relative velocity is zero, and in the latter the ego’s velocity is zero at the point of collision. This approach enhances driver response and safety by providing timely and relevant warnings. Tailored for specific braking dynamics, TTB improves the effectiveness of automotive safety systems.
Singh, Ashutosh PrakashKumawat, HimanshuGupta, Sara
The penetration of ADAS in automotive markets is increasing rapidly. However, their effectiveness and acceptance are significantly influenced by regional driving behaviours and infrastructure. This study explores the interaction between naturalistic driver behaviour in India and the operational characteristics of ADAS systems (FCW, ACC, LCF and BSD) with focus on cars. Using real-world driving data collected from Indian roads, the research aims to highlight the divergence between ADAS design assumptions often based on structured Western traffic environments and the complex, dynamic nature of Indian traffic, characterized by frequent human negotiation, informal road practices, and different vehicle types. The study characterizes multiple driver’s driving pattern through naturalistic driving and ADAS systems behaviour in corresponding situations, notably how they adapt to unstructured Indian scenarios such as lane ambiguity, pedestrian unpredictability, traffic flow unpredictability and frequent road encroachments. The study also aims to identify gaps in ADAS system performance and Indian drivers’ expectations by capturing customer’s voice, underlining the need for context-aware ADAS development tailored to emerging Indian markets. The paper concludes with recommendations for enhancing ADAS usability, safety, and localization strategies in India.
Sankpal, Krishnath NamdevMagar, AkshayKhot, AnkushKulkarni, AlokPerez, Marc
The rapid evolution of intelligent transportation systems has made drivers’ attentiveness and adherence to safety protocols more critical than ever. Traditional monitoring solutions often lack the adaptability to detect subtle behavioral changes in real time. This paper presents an advanced AI-powered Driver Monitoring System designed to continuously assess driver behavior, fatigue, distractions, and emotional state across various driving conditions. By providing real-time alerts and insights to vehicle owners, fleet operators, and safety personnel, the system significantly enhances road safety. The system integrates lightweight AI/ML algorithms, image processing techniques, perception models, and rule-based engines to deliver a comprehensive monitoring solution for multiple transportation modes, including automotive, rail, aerospace, and off-highway vehicles. Optimized for edge devices, the models ensure real-time processing with minimal computational overhead. Alerts are communicated through web and mobile platforms, supplemented by audio-visual cues for prompt user responses. Data from multi-camera setups, auditory sensors, and vehicle CAN bus inputs are processed by a real-time analytics engine that detects abnormal behaviors and safety violations, improving situational awareness and enabling timely interventions. For both individual drivers and fleet managers, the platform serves as an intelligence hub that boosts situational awareness, operational efficiency, and safety compliance. Drivers receive real-time feedback on their behavior, allowing them to make proactive adjustments and reduce risks. Fleet managers can leverage cloud-based connectivity to access predictive analytics, real-time monitoring, and detailed historical behavior data. This enables the identification of unsafe driving patterns, enforcement of safety protocols, and optimization of fleet performance. The system also simplifies regulatory reporting and auditing processes, ensuring compliance with safety standards. By continuously monitoring driver behavior, managers can foster a culture of safety and performance while improving overall fleet operations.
Chikhale, ShraddhaSing, SandipHivarkar, UmeshMardhekar, Amogh
ADAS i.e. Advanced Driver Assistance Systems are pivotal towards amplifying road safety by reducing human error and assisting drivers in critical situations. Most major ADAS technologies are developed and validated using data and test scenarios that are predominantly based on the driving conditions and road environments of developed countries. However, in a country like India, where driving behavior, traffic dynamics, road infrastructure, and accident characteristics differ significantly, the ADAS technologies and test scenarios validated by different forums create a critical gap in deploying such systems on vehicles to work on Indian roads. The major aim of this study was to determine and generate India-specific ADAS test scenarios from the Road Accident Sampling System India (RASSI) database, available MoRTH reports, and data from previously executed ADAS test cases. Through this research, we propose a methodology to identify, extract, and analyze accident scenarios pertaining to the Indian driving environment and their usefulness in developing ADAS technologies to achieve the best results when deployed on vehicles in India. The RASSI database provides exposure to various data factors related to accidents from accident reconstructions, such as velocities at different timestamps, positions of vehicles with respect to time, and pre-crash maneuvers. In addition, the data from MoRTH reports regarding different categories of vehicle collisions will be analyzed extensively for scenario identification and generation for the evaluation of different ADAS functionalities, like Forward Collision Warning, Lane departure warning (LDW), Automatic emergency braking (AEB), and Blind spot detection System. Furthermore, the data generated from the available ADAS Test validations were examined to verify the different parameters related to specific ADAS features. For the purpose of authentication of the pinpointed scenarios, they will be simulated in the simulation software named IPG Carmaker for affirmation of the effectiveness of ADAS functionality if present in the India specific identified accident test scenarios. With the purpose of minimizing the localization gap in ADAS testament, this study provides a data driven, standardized approach towards generation of test scenarios adapted to Indian driving environment.
Adhikari, MayurBhagat, AjinkyaVerma, HarshalKale, Jyoti GaneshKarle, UjjwalaSharma, Chinmaya
Personalized suspension control is pivotal for enhancing vehicle dynamics and ride comfort in intelligent driving systems. This study proposes a driver style recognition model integrating convolutional neural network (CNN) and long–short-term memory (LSTM) networks to match suspension modes with driving styles, validated via a MATLAB–Python co-simulation platform. Time-series multi-source sensor data (throttle position, steering angle, braking intensity) are processed by CNN to extract spatiotemporal features and by LSTM to capture long-term temporal dependencies, enabling accurate classification of aggressive, smooth, and conservative driving styles. A support vector machine (SVM) maps these styles to optimal suspension modes—sport, comfort, or economy—forming an end-to-end framework. Simulation results demonstrate that the CNN–LSTM model achieves an 88% classification accuracy, a 17.33% improvement over the genetic algorithm-optimized backpropagation (GA-BP) model. The SVM-based matching yields matching degrees of 0.95, 0.90, and 0.88 for the three styles, respectively, confirming high accuracy and robustness. Compared to baseline models, the proposed approach excels in prediction accuracy, convergence speed, computational efficiency, generalization, and stability. These findings offer a robust solution for personalized suspension control, enhancing vehicle dynamics and driver comfort.
Wang, ZhuangLiu, JiangSun, HaoyuYuan, YinghaoLiu, JianzeChen, XiaofeiWang, Honglin
The characteristic representation and in-depth understanding of driver personalized driving behavior are fundamental to achieving human-like autonomous driving, enhancing the rationality of autonomous driving decisions, and meeting passengers’ personalized needs. [ADDED]Personalized driving behavior refers to individual-specific patterns in vehicle operation that emerge from drivers’ unique combinations of skills, risk tolerance, and habitual responses.However, current research lacks consideration of cluster analysis in the feature representation stage and ignores the time-varying contribution degree of time series values to low-dimensional features, which inhibits further utilization and development. This study adopts deep embedding clustering method and introduces attention mechanism to investigate driver personalized high-speed lane change behavior.[ADDED] Using a comprehensive driving simulator platform, we collected 15-channel time series data from 12 drivers performing 216 lane changes across 18 controlled scenarios. The research establishes a joint optimization framework that simultaneously learns feature representation and cluster assignment through a variable joint objective function. Results show that compared with baseline methods, the features characterized by this method are closest to the original data after reconstruction. The clustering results demonstrate high intra-cluster compactness, large inter-cluster distances, and clear cluster shapes with significant differences among categories, facilitating personalized driving behavior classification. Model validation on multi-channel temporal classification datasets confirms the efficiency of the proposed model and effectiveness of attention mechanism integration with deep embedding clustering. This study reveals the superiority of attention deep embedding clustering method in the field of driver personalized driving behavior analysis and provides new directions for future research in intelligent vehicle development.
Dong, HaominWang, WeiWang, YueLi, LunYue, YiTian, JiaxiaoHan, Jiayi
Decision modeling based on game theory provides an effective means to achieve safe and efficient ramp merging. However, there are some limitations in the current research, such as previous ramp merge control only studied the interaction problem of networked autonomous vehicles, ignoring the diversity of vehicle types, which is a non-negligible problem in real life. To solve this problem, this study proposes to use different game approaches to address the merging challenge. First, a static game is used to deal with the merging problem of networked self-driving vehicles, and then a belief pool with non-cooperative game approach is used to deal with the problem of human driver’s driving style with the merging problem of self-driving vehicles with human-driven vehicles with unknown information. The simulation results show that the efficiency of on-ramp merging can be significantly improved when networked self-driving cars interact with each other; in the case of merging self-driving cars with human-driven cars, the self-driving cars can recognize the driving styles of the opposite cars and make accurate decisions, which improves the driving efficiency, ensures driving safety and maximizes passenger comfort to the greatest extent.
Gao, ZhenyuDong, JiuyunZhang, LuGuo, Ge
Items per page:
1 – 50 of 762