Calibration of Controlled Markov Chains for Predicting Pedestrian Crossing Behavior Using Multi-objective Genetic Algorithms

Pedestrian motion prediction is a core issue in assisted and automated driving and challenging to solve. In this work, controlled Markov chains are used for predicting pedestrian crossing behavior in urban environments with and without crosswalks. Intentions, such as crossing a road, are estimated by incorporating the probability of colliding with other traffic participants. On a public dataset, we calibrate the model parameters using genetic algorithms which we formulate as a multi-objective optimization problem. Rather than only minimizing the position deviation of the prediction, we also consider the classification performance for pedestrians' crossing intention. The conducted evaluation shows benefits of our approach: it achieves comparable intention recognition performance compared to a support vector machine, while additionally achieving accurate spatiotemporal predictions.