Bridging the Gap: Improving Domain Generalization in Trajectory Prediction

In recent years, there has been a rapid rise in interest in trajectory prediction in the field of autonomous driving. However, the domain generalization of current works is often neglected, and their performance tends to degrade when transferred to a different dataset or scenario. In this article, we present a solution to this problem that accounts for the realistic conditions necessary for autonomous driving. Specifically, we identify velocity and environment as possible causes for the decline in domain generalization. Then, we propose incorporating a module for velocity refinement as a solution to the velocity issue. As a response to the environmental issue, we propose both self-distillation and an environment-specific loss. Our new model is named Lane Transformer++, with one plus representing velocity issues and the other representing environmental concerns. Comprehensive evaluations on both the Argoverse and INTERACTION datasets demonstrate that the proposed method can significantly enhance the performance of prediction.

I. Introduction

Recent years have witnessed a tremendous increase in the development of autonomous driving technology, leading to a great variety of trajectory prediction models being proposed [1], [2], [3]. Trajectory prediction is crucial for the safety of vehicles such as cars, bikes, and pedestrians, as it enables them to anticipate and react to the movements of nearby objects [4], [5]. Furthermore, for traffic systems, the capability to consider future scenarios is essential for dynamic optimization [6], [7], [8]; this can include the timely alteration of traffic lights.