Learning V2V interactive driving patterns at signalized intersections

Semantic understanding of multi-vehicle interaction patterns at intersections play a pivotal role in proper decision-making of autonomous vehicles. This paper presents a flexible framework to automatically extract these interaction patterns from observed temporal sequences based on driving primitives. A Bayesian nonparametric approach is developed to segment vehicle-to-vehicle (V2V) sequential observations into small analytically interpretable components (called driving primitives) without using prior knowledge. Then, an unsupervised clustering method is developed to gather these driving primitives into groups (called driving patterns) by integrating with dynamic time warping. In addition, the extracted driving primitives are used to quantitatively analyze the similarities between behaviors at different intersections using a relative entropy metric, i.e., KullbackLeibler (KL) divergence. Finally, 706 naturalistic V2V events from eight typical urban signalized intersections are used to validate the effectiveness of the proposed primitive-based framework. Experimental results demonstrate that there exist 15 types of interactive driving patterns for V2V behaviors at intersections in our database. Moreover, the distribution of interactive driving patterns could characterize the types of intersections.