Semantically-Aware Pedestrian Intent Prediction With Barrier Functions and Mixed-Integer Quadratic Programming

Abstract

We develop algorithms for inferring long-term intentions and parameters of local collision-avoidance behavior of agents in a multiagent system from their trajectories. This problem is challenging because an agent’s observed trajectory only partially manifests its long-term task; it also contains adjustments made by the agent to ensure collision avoidance with other agents and obstacles in the environment. Since an observer would have no means to determine the magnitude of these adjustments, it is difficult to isolate the task-oriented component from the observed motion. To circumvent this problem, we model the agent’s dynamics using a reactive optimization whose objective function captures the long-term task while its constraints capture collision-avoidance behavior. We develop two robust mixed-integer programming algorithms that infer the task and safety related parameters of this optimization problem from the positions and velocities of the agents. These algorithms are validated on synthetic datasets using parameter estimation errors, displacement errors and computation time as metrics. We further test these algorithms on a dataset of real human trajectories. We show that the learned parameters capture the true underlying pedestrian dynamics by rolling out the learned model and showing similarity between the ground truth trajectories and the reconstructed trajectories.