| With the rapid development of artificial intelligence and hardware and software platform,autopilot,robot and other related technologies have gradually become the focus of the industry.One of the key tasks of secure navigation for autonomous mobile platforms such as self-driving vehicles and social interaction robots is to accurately predict the future trajectories of pedestrians in specific scenarios.However,the complex social interaction between pedestrians and the multi-modal properties of pedestrian movement make pedestrian trajectory prediction a challenging task.To solve these technical problems,this paper proposes an end-to-end multimodal pedestrian trajectory prediction model based on generative adversarial networks and graph attention network.Firstly,the trajectory motion coding of all pedestrians in the scene is taken as the model input,and the graph attention network is introduced to assign different attention weights to each pedestrian,so as to identify the more important pedestrians,accurately model the complex interaction between pedestrians,and obtain the filtered social space interaction information;secondly,the new long short term memory network is used to accurately capture the surrounding lines The continuous time effect of human on the target pedestrian is analyzed,and the spatiotemporal interaction code of the trajectory is obtained.Secondly,the prediction accuracy of the model is further improved by combining the spatio-temporal interaction information and random noise;finally,the performance of trajectory prediction is improved by generating confrontation network to simulate pedestrian distribution,capture the uncertainty of prediction path and generate multiple reasonable future trajectories.The accuracy,inference time and memory utilization of several existing benchmark models and the proposed models are compared on two publicly available data sets.The experimental results show that,compared with the existing pedestrian path prediction methods,the ADE,Average displacement error and FDE,Final displacement error are reduced by 21.9% and 23.8%,respectively.At the same time,the inference time of the model is basically unchanged,and the memory occupancy is only half of the existing prediction model.In conclusion,the method proposed in this paper effectively improves the prediction accuracy of the algorithm while ensuring real-time performance. |
