Verification of Vulnerable Road User Pre-Crash Scenarios in Vision-Based Safety Applications

Verifying training datasets in vision-based vehicle safety applications is crucial to understanding the potential limitations of detection capabilities that may result in a higher safety risk. Vision-based pedestrian safety applications with crash avoidance technologies rely on prompt detection to avoid a crash. This research aims to develop a verification process for vulnerable road user safety applications with vision-based detection functionalities. It consists of reviewing the application’s safety requirements, identifying the target objects of detection in the operational design domain and pre-crash scenarios, and evaluating the safety risks qualitatively by examining the training dataset based on the results of pre-crash scenarios classification. As a demonstration, the process is implemented using open-source pedestrian tracking software, and the pre-crash scenarios are classified based on the trajectories of pedestrians in an example training dataset used in a pedestrian automatic emergency braking system. The dataset provides abundant vulnerable road users under various operating conditions. Results show that the dataset does not cover snowy weather, fog, smoke, smog, and dust/dirt of particulate matter conditions. Crossing, cut-in, and front pre-crash scenarios are less well-represented in the dataset than pedestrians on the sidewalk. Very few pedestrians were found in the front-near zone, which suggests the dataset lacks sufficient imminent pre-crash scenarios that pose the greatest safety risk. The performance of the pedestrian tracking software is evaluated and future research for improving the proposed mechanism is discussed.