A Data-Driven Approach for Simulating Pedestrian Collision Avoidance in Crossroads

This paper presents a model for solving collisions between humans and making them reach their destinations successfully. The proposed model is elaborated from motion capture data. We first adopt dynamic global planning to compute a shortest path to make the pedestrians to get to the other side of the road based on the principle of maximum convenience. In a second stage, we integrate a local collision avoidance algorithm to guarantee that pedestrians do not collide with each other. Our video records prompted us to consider another adaptation - sidestepping besides the reorientation and deceleration strategies. We also consider each pedestrian's individual role in collision avoidance and their adaptation method following nonverbal communication with other pedestrians. Finally we apply real motion capture data to the motions and simulate the road crossing scenario in a 3D virtual environment. The experimental results demonstrate that our approach not only reflects the characteristics of the collision avoidance behaviors well, but also shows more natural and realistic motions of pedestrians compared with RVO and social force model.