SSL-Interactions: Pretext Tasks for Interactive Trajectory Prediction

This paper addresses motion forecasting in multi-agent environments, pivotalfor ensuring safety of autonomous vehicles. Traditional as well as recentdata-driven marginal trajectory prediction methods struggle to properly learnnon-linear agent-to-agent interactions. We present SSL-Interactions thatproposes pretext tasks to enhance interaction modeling for trajectoryprediction. We introduce four interaction-aware pretext tasks to encapsulatevarious aspects of agent interactions: range gap prediction, closest distanceprediction, direction of movement prediction, and type of interactionprediction. We further propose an approach to curate interaction-heavyscenarios from datasets. This curated data has two advantages: it provides astronger learning signal to the interaction model, and facilitates generationof pseudo-labels for interaction-centric pretext tasks. We also propose threenew metrics specifically designed to evaluate predictions in interactivescenes. Our empirical evaluations indicate SSL-Interactions outperformsstate-of-the-art motion forecasting methods quantitatively with up to 8improvement, and qualitatively, for interaction-heavy scenarios.