A Review of Deep Learning-Based Vehicle Motion Prediction for Autonomous Driving

Abstract

Autonomous driving vehicles can effectively improve traffic conditions and promote the development of intelligent transportation systems. An autonomous vehicle can be divided into four parts: environment perception, motion prediction, motion planning, and motion control, among which the motion prediction module plays an essential role in the sustainability of autonomous driving vehicles. Vehicle motion prediction improves autonomous vehicles’ understanding of the surrounding dynamic environment, which reduces the uncertainty in the decision-making system and facilitates the implementation of an active braking system for autonomous vehicles. Currently, deep learning-based methods have become prevalent in this field as they can efficiently process complex scene information and achieve long-term prediction. These methods often follow a similar paradigm: encoding scene input to obtain the context feature, then decoding the context feature to output predictions. Recent research has proposed innovative improvement designs to enhance the primary paradigm. Thus, we review recent works based on their improvement designs and summarize them based on three criteria: scene input representation, context refinement, and prediction rationality improvement. Although most works focus on trajectory prediction, this paper also discusses new occupancy flow prediction methods. Additionally, this paper outlines commonly used datasets, evaluation metrics, and potential research directions.