基于层次图注意的异构多目标轨迹预测方法

Heterogeneous Multi-object Trajectory Prediction Method Based on Hierarchical Graph Attention

有效预测周边多目标的未来轨迹是智能汽车决策和运动规划成功的关键。大多数现有研究考虑车辆个体行为之间的成对交互关系,而忽略异构交通参与者之间不同的反应模式和其他场景因素对预测的影响,使得预测轨迹的合理性低,影响运动控制的安全性。鉴于此,本文提出了一种基于层次图注意的异构多目标轨迹预测方法HGATP,首先创新性地构建类别-目标-车道的3层次图,并使用GRU和GCN分别对不同类型的目标进行独立的类别编码,捕捉不同类别的特征;其次,为强化异构目标交互图的边缘表示,构建层次图注意力机制分别获取类别和类别之间的交互以及目标和车道之间的交互,实现异构多目标间高效交互和共享地图;最后,基于目标轨迹信息和区域的车道信息构建预测网络预测多目标的轨迹。为评估模型性能,分别在INTERACTION和nuScenes数据集上进行实验。实验表明,所提模型在nuScenes数据集上单目标轨迹输出的平均误差损失和最终位移损失均减小了20%以上,在INTERACTION数据集上多目标轨迹输出的ADE损失效果较基线方法减小了2 m误差,提升了复杂道路结构下车辆行驶轨迹预测的合理性。

Effectively predicting the future trajectories for surrounding multiple targets is critical to the suc⁃cess of autonomous vehicle decision-making and motion planning.Most existing studies consider pairwise interac⁃tions between individual vehicle behaviors,while ignoring the influence of different reaction patterns among hetero⁃geneous traffic participants and other scene factors on prediction,which reduces the rationality of the predicted tra⁃jectories and affects the safety of motion control.In view of this,this paper proposes a heterogeneous multi-target tra⁃jectory prediction method HGATP based on hierarchical graph attention.Firstly,a category-target-lane three-level graph is innovatively constructed,and different types of targets are independently coded with categories using GRU and GCN respectively to capture the features of different categories.Secondly,to enhance the edge representation of the heterogeneous target interaction graph,the attention mechanism of hierarchical graph is constructed to separate⁃ly capture the interaction between categories and categories and the interaction between targets and lanes so as to achieve efficient interaction and sharing of maps among heterogeneous multiple targets.Finally,a prediction net⁃work is constructed to predict the trajectories of multiple targets based on the target trajectory information and the lane information of the region.To evaluate the performance of the model,experiments are conducted on the INTER⁃ACTION and nuScenes datasets respectively.The experiments show that the proposed model reduces the average displacement error and final displacement error of single-target trajectory output on the nuScenes dataset by more than 20%,with the ADE loss effect of multi-target trajectory output on the INTERACTION dataset reduced by 2 m error compared with the baseline method,which improves the reasonableness of vehicle trajectory prediction under complex road structures.