A Safe Driving Decision-Making Methodology Based on Cascade Imitation Learning Network for Automated Commercial Vehicles

Safe driving decision-making is particularly important for automated commercial vehicles (ACVs). Small passenger vehicles pay more attention to collision prevention, while commercial vehicles with a longer brake distance and worse roll stability need to consider both anti-collision and anti-rollover. The widely studied driving decision-making methods for small passenger vehicles cannot be simply and directly applied to commercial vehicles. This article proposes a safe driving decision-making methodology based on a cascade imitation learning network (CILN). The CILN integrates two parts, namely, the supervised learning part and the imitation learning part. The first part learns safe driving maneuvers extracted from naturalistic vehicle sensor data. Through sensor data processing, it develops decision-making at a humanoid level, such as avoiding jerky driving actions. In the second part, generative adversarial imitation learning (GAIL) is introduced to further learn safe driving decisions under conditions prone to collision and rollover. Finally, both highD dataset and simulation of urban mobility (SUMO) are used to train and verify the performance of the CILN. By comparing the evaluation indicators of time to collision (TTC), reverse TTC (RTTC), deceleration rate to avoid the crash (DRAC), distance headway, and lateral acceleration, the CILN outperforms the other decision-making algorithms. Experimental results show that the CILN can provide safe driving decision-making and ensure the driving safety of ACVs in dense traffic flow.

Automated summary

Safe driving decision-making is crucial for automated commercial vehicles (ACVs). Unlike small passenger vehicles, commercial vehicles need to consider both collision prevention and rollover prevention due to their longer brake distance and worse roll stability. This article proposes a safe driving decision-making methodology based on a cascade imitation learning network (CILN), which integrates supervised and imitation learning parts. Experimental results show that the CILN outperforms other decision-making algorithms, ensuring the driving safety of ACVs in dense traffic flow.