基于时空图的行人多模态轨迹预测方法

Pedestrian multi-modal trajectory prediction method based on spatial-temporal graph

为提高行人轨迹预测的准确率与速度,提出一种基于时空图的GAN轨迹预测模型。使用LSTM节点将人与环境交互的时空图转换为前馈可微分特征编码,提出用于集成场景上下文信息的全局节点,通过缩放点积注意力机制研究全局交互对行人轨迹的相对影响,使用生成对抗网络体系结构对模型在数据集中进行训练。此外,为使预测结果更加合理,结合行人历史轨迹运用控制点生成一组行人未来轨迹曲线作为假设建议,并与以上预测结果进行相似度分析。实验结果表明,所提模型预测速度与精准度均优于其它方法。

To improve the accuracy and speed of pedestrian trajectory prediction,a GAN trajectory prediction model based on spatial-temporal graph was proposed.The LSTM node was used to convert the spatial-temporal graph of human-environment interaction into feed-forward differentiable feature coding,and a global node for integrating scene context information was innovatively proposed.The relative influence of global interaction on pedestrian trajectory was studied through the sealed dot product attention mechanism.Gencrativc adversarial network architecture was used to train the model in the data set.To make the prediction results more reasonable,a set of pedestrian future trajectory curves generated by combining the historical trajectory of pedestrians with control points were used as hypothetical suggestions,and similarity analysis was carried out with the above prediction results.Experimental results show that the prediction speed and accuracy of the proposed model are better than that of other methods.