Highly Interactive Self-Supervised Learning for Multi-Modal Trajectory Prediction

Abstract

To ensure the safety of autonomous vehicles, trajectory prediction is critical as it enables vehicles to anticipate the movements of surrounding agents, thereby facilitating the planning of secure and strategic driving routes. However, striking a trade-of between predictive accuracy and training costs has always been an intricate challenge. This paper introduces a groundbreaking framework for trajectory prediction known as Highly Interactive Self-Supervised Learning (HI-SSL), a methodology based on self-supervised learning (SSL) that has yet to be thoroughly investigated in the realm of trajectory prediction. The cornerstone of the aforementioned framework is Interactive Masking, which leverages a novel trajectory masking strategy facilitating self-supervised learning tasks that not only enhance prediction accuracy but also eliminate the need for manual annotations. Experiments conducted on the Argoverse motion forecasting dataset demonstrate that our approach achieves competitive performance to prior methods that depend on supervised learning without additional annotation costs.