Deep Learning Based Trajectory Prediction in Autonomous Driving Tasks: A Survey

Abstract

Nowadays, autonomous driving technology has greatly advanced. Trajectory prediction, as a significant part of autonomous driving, is highly valued because it serves as a connecting link between the tracking module and the planning module. Therefore, it is of great value to review the existing trajectory prediction methods and summarize their advantages and disadvantages. This paper provides a thorough review of trajectory prediction methods. Specifically, based on the types of input information, we classify trajectory prediction methods into those using temporal information only and those combining both temporal and spatial information. For each category, we comprehensively introduce specific methods along with associated challenges, shortcomings, and research gaps. Finally, we summarize the advantages and disadvantages of these methods and compare their performance on public datasets for future research in trajectory prediction.