基于时空融合LSTM网络的驾驶视角轨迹预测

自动驾驶环境感知系统的重要任务之一是对周围交通目标进行轨迹预测，其输出轨迹可为决策控制和路径规划提供所需目标信息。考虑传统轨迹预测方法一般基于俯视视角而难以满足自动驾驶车载感知的实际需求，提出一种基于长短时记忆(LSTM)网络模块、空间交互模块和时间行为注意力模块相融合的驾驶视角轨迹预测算法。为更好体现交通目标与周围环境的交互作用与不确定性，将交通目标建模为道路智能体。在LSTM网络模块，利用状态增强LSTM算法对于单智能体信息结合周围邻居隐藏状态信息对轨迹历史数据进行挖掘，输出基准轨迹预测结果。在空间交互模块，对智能体进行图建模并采用图空间交互方法对当前智能体与周围智能体空间交互分析。在时间注意力模块，对智能体驾驶行为进行细粒度分类并采用时间注意力估计智能体驾驶行为影响。最后，利用以上时空模块对原始轨迹结果进行增强，获取最终增强后驾驶视角交通智能体轨迹。为验证所提出算法的有效性，在符合中国复杂交通特点的D²-City行车记录数据集上训练、测试和验证算法，并与标准LSTM、Social LSTM、Social GAN等多种国际领先的轨迹预测算法进行定量和定性对比分析。研究结果表明:所提出算法能够在不同输入和预测时长下都取得有竞争力的结果，全局位移误差指标平均比其余3种算法降低超过20%，能够显著提高轨迹预测精度，更符合自动驾驶环境感知需求。

Driver Perspective Trajectory Prediction Based on Spatiotemporal Fusion LSTM Network

One of the most important tasks in the autonomous driving environment perception system is to predict the trajectories of surrounding traffic objects, which the output trajectories can provide information for vehicle control, decision-making, and path planning. Considering traditional trajectory prediction usually based on bird-view trajectory prediction，which couldn't satisfy real demand of onboard autonomous driving environment perception. In this paper, we proposed a novel driver perspective trajectory prediction algorithm, which combines of LSTM (Long Short-Term Memory) module, spatial interaction module, and temporal behavior attention module. In order to reflect the interaction and uncertainty between traffic objects and surrounding environment, we modeled the traffic objects as traffic agents. As LSTM network module, the SR-LSTM method was employed to combine single agent's information with surrounding neighbor agents' hidden information to mine historical trajectory information. As spatial interaction module, graph modeling was performed on the agents, and the spatial interactions between the current agent and surrounding agents was analyzed, which using the graph space interaction method. For temporal behavior attention, the driving behaviors of agents was classified into fine-grained categories, the influence of temporal attention on the agent's driving behavior was estimated. Finally, the above spatial module and temporal module were used to enhance the original trajectory results to output final refinement trajectories of traffic agents under driver perspective. To verify the effectiveness of proposed method, we operated our algorithm on the D²-City driving recorder benchmark that can reflect China's complex traffic characteristics, and performed quantitative and qualitative comparative analysis with competitive trajectory prediction algorithms: vanilla LSTM, Social LSTM, and Social GAN. The research results show that the algorithm proposed in this paper can achieve competitive results under different input and prediction durations. In comparison with other three methods, the proposed algorithm's final displacement error index is reduced more than 20% on average. The proposed method can significantly improve the trajectory prediction accuracy, which is suitable for automatic driving environment perception.