Identifying Pedestrian-Vehicle Interaction Patterns at Unsignalized Intersection Using Social LSTM and Dynamic Time Warping

The uncertainty and randomness of pedestrian behavior at unsignalized intersections make the decision process of autonomous vehicles difficult. Therefore, this study aims to investigate the interaction patterns between pedestrians and vehicles by capturing the changes. To this end, one parameter system and an interaction-identifying framework were established by analyzing the DUT Data set containing 816 trajectories. Firstly, the Social LSTM is utilized to extract interaction information from the motion features and participant types. Subsequently, the k-means algorithm based on Dynamic Time Warping (DTW) is applied to obtain the behavior patterns that make up the crossing process. The results indicate that the Social LSTM outperforms conventional LSTM in extracting interaction information. Meanwhile, the four obtained typical interaction clusters are truly valid and effective. Overall, this study can provide a new insight into establishing a scenario library of interaction behavior patterns for autonomous driving.