Personalized Driver Model for On-Ramp Merging Using LSTM

2026-01-0525

04/07/2025

Authors
Abstract
Content
This study develops a personalized driver model for freeway merging, embedding individual driving characteristics into automated longitudinal and lateral control through Long Short-Term Memory (LSTM) networks. Uniform assistance strategies such as adaptive cruise control (ACC) can feel uncomfortable when misaligned with a driver's style. To address this, we focus on the merging maneuver, a safety-critical task requiring anticipation and timing, and examine how roadway and environmental information contributes to model fidelity. Driving data were collected in a high-fidelity motion-base simulator replicating merging sections. In Scenario 1 (Tokyo-Nagoya Expressway, three lanes), seven licensed male drivers in their twenties each completed 15 trials. In Scenario 2 (Metropolitan Expressway, two lanes, lower speed), six drivers followed the same protocol. Model inputs included ego speed, relative distance and speed to the lead vehicle, distances to the end of the acceleration lane and to the hard nose, and in Scenario 2 a time-headway-like measure as distance to end normalized by ego speed. Outputs were ego longitudinal acceleration and both longitudinal and lateral accelerations. Each model was trained per driver and evaluated by root mean squared error (RMSE), with 0.5 m/s² the threshold for high accuracy. Comparing input patterns showed that incorporating merging-environment variables (Pattern 1) reduced errors compared with excluding them (Pattern 2): average RMSE was 0.29 versus 0.41 m/s² across drivers (Mann-Whitney test, p <0.05). Extending the output to lateral acceleration captured merge timing while maintaining precision; average RMSEs were 0.071 (longitudinal) and 0.016 (lateral) in Scenario 1, and 0.085 and 0.020 in Scenario 2, all well below 0.5 m/s². A playback experiment and questionnaire assessed subjective impressions: discomfort, perceived safety, timing, similarity to one's driving, and willingness to use such models. Limitations remain, including verifying performance across roadway geometries and speed ranges, to be addressed in future work.
Meta TagsDetails
Citation
Shen, Shuncong and Toshiya Hirose, "Personalized Driver Model for On-Ramp Merging Using LSTM," SAE Technical Paper 2026-01-0525, 2025-, .
Additional Details
Publisher
Published
Apr 7, 2025
Product Code
2026-01-0525
Content Type
Technical Paper
Language
English