Can automated driving prevent crashes with distracted Pedestrians? An exploration of motion planning at unsignalized Mid-block crosswalks

• Automated vehicles’ motion planning models for preventing crashes with distracted pedestrians on unsignalized mid-block crosswalks are proposed. • Rule-based model leads to 5.1% serious conflicts and 1.4% crashes. • Learning-based and hybrid models are based on policy gradient reinforcement learning. • Learning-based model has good safety performance but is overly sensitive to risks. • Hybrid model produces reasonable and safe trajectories with acceptable efficiency loss. Pedestrian distraction may provoke severe difficulties in automated vehicle (AV) control, which may significantly affect the safety performance of AVs, especially at unsignalized mid-block crosswalks (UMCs). However, there is no available motion-planning model for AVs that considers the effect of pedestrian distraction on UMCs. This study aims to explore innovative approaches for safe and reasonable automated driving in response to distracted pedestrians with various speed profiles at UMCs. Based on two common model design concepts, two new models are established for AVs: a rule-based model that solves motion plans through a fixed calculation procedure incorporating several optimization models, and a learning-based model that replaces the deterministic optimization process with policy-gradient reinforcement learning. The developed models were assessed through simulation experiments in which pedestrian speed profiles were defined using empirical data from field surveys. The results reveal that the learning-based model has outstanding safety performance, whereas the rule-based model leads to remarkable safety problems. For distracted pedestrians with significant crossing-speed changes, rule-based AVs lead to a 5.1% probability of serious conflict and a 1.4% crash probability. The learning-based model is oversensitive to risk and always induces high braking rates, which results in unnecessary efficiency loss. To overcome this, a hybrid model based on the learning-based model was developed, which introduces a rule-based acceleration value to regularize the action space of the proposed learning-based model. The results indicate that the hybrid approach outperforms the other two models in preventing crash hazards from distracted pedestrians by employing appropriate braking behaviors. The high safety performance of the hybrid models can be attributed to the spontaneous slowing down of the vehicle that initiates before detecting pedestrians on UMCs. Although such a cautious driving pattern leads to extra delay, the time cost of the hybrid model is acceptable considering the significant improvements in ensuring pedestrian safety.