Addressing Mode Collapse in Trajectory Prediction: A Maneuver-Oriented Metric and Approach

Latest advancements in deep learning architectures and the availability of several open source trajectory prediction benchmarks have given a great boost in the trajectory prediction research applied to the autonomous driving domain. Most advanced trajectory prediction methods focus on minimizing the prediction error in terms of expected displacement error. Although such methods, for a given traffic participant, will propose multiple future trajectories, several of the high probability trajectories will correspond to slightly different realizations of the most probable high level maneuver such as turning left, slowing down, etc.For this reason, in this paper we aim to evaluate SOTA trajectory prediction models by examining their ability to predict diverse future maneuvers with different semantic meaning. To accomplish this, we propose a new evaluation metric called Maneuver Miss Rate (MMR). Furthermore, we present a Maneuver Oriented Trajectory Prediction model (MOTP) that tackles the trajectory prediction problem from a maneuver perspective. We demonstrate that MOTP is capable of proposing multiple trajectories, each one describing a different high level behavior. This approach provides us with direct control over trajectory diversity which results in a reduction of our proposed MMR metric when compared to our SOTA baselines on the Argoverse dataset.