| With the development of intelligent interconnection and 5G technology,unmanned driving and intelligent transportation have been fully applied,and researchers are gradually deepening their research in this field.Pedestrian trajectory prediction,as an important part of it,has become a hot research direction.At present,pedestrian trajectory prediction still faces problems such as insufficient accuracy and large errors.There are two main reasons for this.One is that due to the randomness of pedestrians in the walking process,it is difficult to model the positional interaction of pedestrians in the same space and the higher-level interaction between a single pedestrian and global pedestrians.The second is that the current main work is based on the data-driven method,which has a strong data fitting ability,but there are still problems of insufficient accuracy and lack of interpretability.In view of the above problems,the research is carried out from the following two aspects:(1)Consider modeling the position interaction information of pedestrians in the same space and the higher-level interaction between individual pedestrians and global pedestrians.A multi-head soft attention graph convolutional network is proposed.Firstly,the interaction and movement trend of pedestrians are modeled from the spatial graph and the temporal graph to obtain a sparse spatial directed graph and a sparse temporal directed graph,and then through the graph convolutional network The trajectory features are learned,and finally the feature information is processed through the time convolutional network to obtain the double Gaussian distribution parameters,that is,the predicted trajectory of pedestrians.(2)Aiming at the problems of interpretability of pedestrian interaction and insufficient accuracy of trajectory prediction in the process of pedestrian movement,a pedestrian force model combined with neural network,namely improved avoidance force algorithm,is proposed.The algorithm first generates endpoint information given the observed pedestrian trajectory,and then completes the entire trajectory by improving the avoidance force,and finally optimizes the improved avoidance force trajectory through the trajectory optimization network,refines and refines the pedestrian’s predicted trajectory.The experimental results on the public datasets show that the proposed multi-head soft attention graph convolutional network method can better construct the interaction between pedestrians,and has a great improvement in the performance of pedestrian trajectory prediction.The displacement error ADE and final displacement error FDE are reduced by 2.8% and 16.92%respectively.The proposed improved avoidance force-based algorithm correctly models the interaction and avoidance patterns between pedestrians,improves the pedestrian interaction interpretability problem,and reduces the average displacement error ADE and final displacement error FDE metrics by 6% and 16% for the data sets ETF and UCY,respectively. |
