基于几何约束稀疏图的行人多模态轨迹预测方法OA

Aiming at the problems of inaccurate interaction modeling and insufficient representation of multimodal uncertainty in pedestrian trajectory prediction for complex dynamic scenes,this study proposed a pedestrian multimodal trajectory predic-tion method named GCSG-Traj based on geometric constrained sparse graphs.This method constructed a sparse interaction graph under three constraints including viewing scope,distance threshold and motion approaching property,extracted spatio-temporal features by integrating graph convolutional networks and Transformers,and realized multimodal trajectory prediction through two-dimensional Gaussian distribution parameter prediction and Latin hypercube sampling.Experimental results on the ETH/UCY datasets show that GCSG-Traj achieves an average displacement error of 0.31 m and a final displacement error of 0.51 m.It outperforms most comparison methods and maintains an excellent balance between prediction accuracy and model compactness.The results verify that this method effectively improves the accuracy and diversity of pedestrian trajectory predic-tion in complex dynamic scenes.