A Survey of Deep Learning-Based Multimodal Vehicle Trajectory Prediction Methods

With the rapid development of deep learning, an increasing number of researchers have paid attention to and engaged in research in the field of autonomous driving, achieving numerous remarkable achievements. The paper aims to explore the current status and methods of trajectory prediction for autonomous vehicles, and reviews some commonly used trajectory prediction techniques and algorithms. The existing vehicle trajectory prediction tasks mainly involve extracting and modeling the historical trajectory sequences of the target vehicle and the surrounding environmental information features, to infer multimodal trajectories for a certain future time duration. The paper focuses on deep learning-based multimodal trajectory prediction methods, including traditional convolutional neural networks, recurrent neural networks, graph neural networks, attention mechanisms, and the latest research results of mixed models composed of multiple network structures. Additionally, two large-scale publicly available datasets in the field of autonomous driving, the Argoverse dataset and the nuScenes dataset are introduced, and the performance of existing evaluation metrics is analyzed. Finally, through the analysis and comparison of these methods, the paper summarizes and deduces the possible future directions of vehicle trajectory prediction and the existing technological bottlenecks, aiming to provide useful assistance and prospects for future research and application of autonomous vehicle trajectory prediction.