AI 驱动的自动驾驶汽车轨迹预测方法综述

In autonomous driving systems, trajectory prediction plays an important role in connecting vehicle's perception and decision-making, and enhancing driving safety and overall system robustness. In recent years, with the continuous advancement of artificial intelligence (AI), the AI-driven trajectory prediction methods have seen significant progress in terms of accuracy, adaptability, and the capability to model complex traffic environment. This paper provides a systematic review of mainstream trajectory prediction methods in autonomous driving, with a focus on predictive model frameworks. It first revisits traditional physics-based approaches, and then highlights current research trends, including modeling paradigms based on classical machine learning, deep neural networks, and reinforcement learning. Additionally, recent developments in explainable AI techniques aimed at improving model transparency and safety are discussed. Based on comparative analysis, the paper evaluates the strengths and limitations of various models in interaction modeling, multimodal uncertainty, and generalization capability. Furthermore, it organizes trajectory prediction evaluation metrics and publicly available trajectory prediction datasets according to their characteristics and application scenarios, and summarizes representative real-world deployments from both domestic and international sources. At last, considering the existing research bottlenecks and future development trends, the paper outlines potential directions for future studies, such as enhancing model interpretability, effectively integrating multimodal information, and designing unified frameworks for joint prediction and planning. The purpose of this review is to provide insights and references that can be used in the future research and applications.