融合距离阈值和双向TCN的时空注意力行人轨迹预测模型

Fusion of distance threshold and Bi-TCN for spatio-temporal attentionpedestrian trajectory prediction model

为解决因缺乏部分行人建模思想、缺少时间维度的全局视野和忽略行人交互模式多样性,而导致交互建模不充分、低预测精度等问题,提出基于Social-STGCNN(social spatio-temporal graph convolutional neural network)的改进模型STG-DTBTA(spatio-temporal graph distance threshold Bi-TCN attention)。首先,构建PPM(partial pedestrian module)模块,对不满足距离阈值等约束条件的行人交互连接剪枝以去噪。其次,引入时空注意力机制,空间注意力动态分配交互权重,并设置多个注意力头以处理交互多样性问题;时间注意力捕捉时序数据的时间依赖关系。最后,采用双向TCN增加全局视野以捕捉轨迹数据中的动态模式和趋势,并采用门控机制融合双向特征。在ETH和UCY数据集上的实验结果表明,与Social-STGCNN相比,STG-DTBTA在维持参数量与推理时间接近的情况下,ADE平均降低8%,FDE平均降低16%。STG-DTBTA具有良好的交互建模能力、模型性能和预测效果。

In order to solve the problems such as insufficient interaction modeling and low prediction accuracy due to the lack of partial pedestrian modeling ideas,the lack of global vision in time dimension,and the neglect of the diversity of pedestrian interaction modes,this paper proposed an improved model STG-DTBTA based on Social-STGCNN.Firstly,the model constructed PPM module,and pruned the pedestrians inter links that were not meet constraints such as distance threshold for de-noising.Secondly,the model introduced the spatio-temporal attention mechanism.Spatial attention dynamically assigned interactive weights,and set up multiple attention heads to deal with the interaction diversity problem.Temporal attention captured temporal dependencies of temporal data.Finally,the model used Bi-TCN to increase global perspective to capture dynamic patterns and trends in trajectory data,and used gating mechanism to incorporate the bidirectional features.The experimental results on the datasets ETH and UCY show that compared with Social-STGCNN,ADE and FDE are decreased by an ave-rage of 8%and 16%respectively when the number of parameters and the inference time kept close to it.The STG-DTBTA has good interactive modeling ability,model performance and prediction effect.