基于深度神经网络的轨迹预测算法综述

A Survey of Trajectory Prediction Algorithms Based on Deep Neural Networks

人类驾驶者会持续观察、分析周边车辆和行人的行为,实时地规划安全的行车轨迹。自动驾驶汽车也应当与人类一样具备感知和预测交通参与者行为的能力,以提前判断其未来的运动轨迹。轨迹预测模块的预测准确度至关重要,因为其接受感知系统提供的输入信息,并作为路径规划等决策任务的上游输入,使其成为自动驾驶技术中承上启下的重要中间环节。随着近年来数据科学和传感器领域的长足发展,大量关注行人、车辆等多样化交通参与者的大型数据集得以建立,使得轨迹预测问题的解决方案从传统的动力学模型过渡到深度学习模型成为可能。基于此,介绍行人、车辆轨迹预测算法的发展历程和重要论文的解决方案,并总结该领域形成共识的几种思路,展望最新的研究趋势。

Human drivers will observe and analyze the behavior of surrounding traffic agents to plan safe driving trajectories in realtime. Self-driving cars should also have the same ability as humans to perceive and predict the behavior of traffic participants to judge their future trajectory in advance. The trajectory prediction module accepts the input information provided by the perception system and serves as the upstream input for decision-making tasks such as path planning. This makes the module an important intermediate link in the automatic driving technology, and its accuracy is crucial. With the rapid development of data science and sensor technology in recent years, a large number of datasets focusing on diversified traffic participants such as pedestrians and vehicles have been established, making the transition from the traditional dynamic model to the deep learning model possible. This article will introduce the development history of pedestrian and vehicle trajectory prediction algorithms and the solutions of important papers. In addition, the paper will summarize several ideas that have reached a consensus in this field and look forward to the latest research trends.