Self-Supervised Transformer for Trajectory Prediction Using Noise Imputed Past Trajectory

Abstract

Trajectory prediction is one of the important components for achieving higher levels of Society of Automotive Engineers (SAE) driving automation, enabling them to navigate through complex driving scenarios and make informed decisions in unexplored roads. Problems such as the varying behaviour of drivers on the road, sensor measurement inaccuracies, and dense and complex environmental circumstances make this task difficult. To address these challenges, a self-supervised transformer (SST) framework is proposed. The noise-imputed trajectory points for road agents are generated. This enhances the model’s ability to handle uncertain data. A self-supervised task is proposed, which focuses on predicting past trajectory points using both true and noise-imputed trajectory encoded features. This approach highlights the important patterns or connections in the data that could go unnoticed if supervised tasks were the only one used. Further, in addition to the trajectory prediction task, a consistent loss function should be introduced to preserve spatial consistency with noise imputed trajectory points. Moreover, learnable query embeddings are added to the system to improve the diversity of multi-modal predictions. The SST model has been evaluated on the widely used and recent Argoverse 2 dataset and outperforms state-of-the-art models by a margin of 2.17%-24.15%, 5.38%-30.05%, 9.41%-46.89% and 0.20%-21.21% on the minADE6, minFDE6, MR6 and b-FDE6 respectively.