基于异构图学习的交通场景运动目标感知

为了提高无人车在交通场景中的运行效率和运输安全，研究了基于异构图学习的交通场景运动目标感知; 考虑实际交通场景中运动目标之间的复杂交互关系对目标运动的影响，基于异构图学习提出了交通场景中多目标检测-跟踪-预测一体化感知框架; 结合YOLOv5和DeepSORT检测并跟踪运动目标，获得目标的运动轨迹; 使用长短期记忆(LSTM)网络从目标历史轨迹中学习目标的运动信息，使用异构图学习目标间的交互信息，以提高运动目标轨迹预测准确度; 使用LSTM网络对目标运动信息与交互信息解码得到目标未来轨迹; 为了验证方法的有效性，在公共交通数据集Argoverse、Apollo和NuScenes上进行了评估。分析结果表明:结合YOLOv5和DeepSORT可实现对运动目标的检测跟踪，对交通场景中的运动目标实现了75.4%的正确检测率和61.4%的连续跟踪率; 异构图能够有效捕捉运动目标之间复杂的交互关系，并且捕捉的交互关系能够提高轨迹预测精度，加入异构图捕捉交互关系后，运动目标的平均位移预测误差降低了63.0%。可见，考虑交通场景中运动目标之间的交互关系是有效的，引入异构图学习运动目标之间的交互关系可以感知运动目标的历史与未来运动信息，从而帮助无人车更好地理解复杂交通场景。 

Perception of moving objects in traffic scenes based on heterogeneous graph learning

In order to improve the operation efficiency and transportation safety of unmanned vehicles in traffic scenes, the perception of moving objects in traffic scenes was investigated based on the heterogeneous graph learning.In view of the influence of complex interaction relations between moving objects on their motions in actual traffic scenes, an integrated perception framework of multi-object detection-tracking-prediction was proposed based on the heterogeneous graph learning. YOLOv5 and DeepSORT were combined to detect and track the moving objects, and the trajectories of the objects were obtained. The long short-term memory (LSTM) network was used to learn the objects' motion information from their historical trajectories, and a heterogeneous graph was introduced to learn the interaction information between the objects and improve the prediction accuracies of the trajectories of moving objects. The LSTM network was also utilized to decode the objects' motion and interaction information to obtain their future trajectories, and the method was evaluated on the public transportation datasets Argoverse, Apollo, and NuScenes to verify its effectiveness.Analysis results show that the combination of YOLOv5 and DeepSORT can realize the detection and tracking of moving objects and achieve a detection accuracy rate of 75.4% and a continuous tracking rate of 61.4% for moving objects in traffic scenes. The heterogeneous graph can effectively capture the complex interaction relations between moving objects, and the captured interaction relations can improve the accuracy of trajectory prediction. The error of the predicted average displacement of moving objects reduces by 63.0% after the interaction relations captured by the heterogeneous graph are added. As a result, it is effective to consider the interaction relations between moving objects in traffic scenes. The historical and future motion information of moving objects can be perceived by introducing the heterogeneous graph to capture the interaction relations between moving objects, so as to facilitate unmanned vehicles to better understand complex traffic scenes. 4 tabs, 9 figs, 36 refs.