IDSSL: Intent Detection Module Based on Self-Supervised Learning for Trajectory Prediction

Abstract

Trajectory prediction models play a crucial role in various applications such as autonomous driving and human–robot interaction. For trajectory prediction models, understanding and accounting for the intentions of vehicles, such as whether they intend to change lanes, make turns, or maintain their current course, is vital for enhancing the safety and efficiency of autonomous systems. However, one significant limitation of some existing trajectory prediction models is their failure to fully consider the impact of vehicle intentions on agent trajectories. This paper addresses the problem that some existing trajectory prediction models do not fully consider the impact of vehicle intentions on agent trajectories. To tackle this challenge, we propose a self-supervised learning-based module for detecting vehicle intentions, which can be added to most trajectory prediction models. The module consists of a simple and effective pretext task and a lightweight intent detection model. We conduct extensive experiments on a large-scale motion forecasting benchmark to demonstrate the effectiveness of the proposed approach in improving the performance of trajectory prediction baselines.