A Data-driven Markovian Framework for Multi-agent Pedestrian Collision Risk Prediction

Paper in proceeding, 2019

Efficient travel with an Autonomus Vehicle (AV) through urban traffic scenes requires the AV to plan its path as a function of the risk of collision with nearby road users. We propose a framework based on Markov Chains to model the joint behavior of multiple road users, learn their joint behavior from previously observed behaviours, and predict the collision risk based on the learnt model. Using our framework we formulate two models which are based on position only and both position and velocity, respectively. The learning relies on real measurements of pedestrians at a crosswalk. Comparing the models on prediction horizons up to 6 s, we find that including velocity significantly changes the collision risk and that the model better predicts the pedestrian's future position. © 2019 IEEE.