Multi-agent Trajectory Prediction Based On Generative Adversarial Network

Behavior analysis is a hot topic in the field of computer vision in recent years.It has attracted extensive attention from both domestic and overseas scholars and can be applied to many fields such as visual surveillance,content-based retrieval,and intelligent transportation.In this paper,representative trajectory prediction algorithms are analyzed.To overcome the shortcomings of existing algorithms,we proposed an effective agent trajectory prediction algorithm.The main work is as follows:(1)The traditional trajectory prediction algorithm only depends on the motion information of other pedestrians in the scene,and does not consider the dynamic change of the environment within the scene.In this paper,we propose a new algorithm that combines spatial attention and temporal attention.Spatial attention maps the original image sequence through convolutional neural network output features,and gets the weight representation of spatial attention by soft max.After weighted operation,spatial attention is obtained.The temporal attention first compute Euclidean distances between all neighbor agents and target agents in the scene are calculated as input of embedded layer and multi-layer perceptron,and the output is processed by extremum,only considering the information which has great influence on the path planning of target agents.The attention algorithm proposed in this paper pays attention to the dynamic changes of pedestrian motion information and environment in the scene at the same time,so that the target agent can plan the path according to its geographical environment and the movement trend of other pedestrians in the scene.(2)The existing trajectory prediction model only outputs average trajectory,which can not simulate the uncertainty of pedestrian motion well.In this paper,a multi-mode trajectory prediction algorithm based on generating adversarial network is proposed.We use simulation data to test the generation adversarial network and the variable auto-encoder,respectively,to verify whether the generation model can get multi-mode output.According to the simulation results,the generation adversarial network is regarded as the basic framework of the trajectory prediction model,and the generation mechanism of the generation adversarial network is based on the unique generation mechanism of the generation adversarial network.The model can output many reasonable trajectories.(3)We validate the algorithm proposed on the standard datasets ETH,UCY and compared with the state-of-the-art results.Experimental results demonstrate that the proposed multi-modal trajectory model can accurately predict the movement trend of pedestrians in the future.