Research On Pedestrian Trajectory Prediction Methods In Real-World Scenarios

The interaction between artificial intelligence creations and humans is based on the understanding of human behavior and intentions.The pedestrian trajectory prediction task,based on past environments,behavior,and historical trajectories,can help us better predict future trends.This technology has been widely used in population density control,traffic management,service robots,and the increasingly popular field of autonomous driving,with significant research value and significance.In recent years,the research on pedestrian trajectory prediction tasks has increased,but due to the uncertainty of pedestrian intentions and the complexity of interactions in complex real-world scenarios,pedestrian trajectory prediction is still a highly challenging task.This thesis focuses on the research of pedestrian trajectory prediction methods in real-world scenarios,addressing the problem of low prediction accuracy and complex prediction models in most current research.The main research contents are as follows:(1)To address the problem that existing models use a large amount of computing power to consider the interaction of the target pedestrian,making it difficult to perform trajectory inference tasks within the running time in real-world scenarios,a pedestrian trajectory prediction model that integrates trajectory and visual features is proposed to balance inference efficiency and prediction accuracy.The model aims to have as few parameters as possible while fully utilizing the input information.First,the historical trajectory and visual information of the target pedestrian are extracted for feature extraction,obtaining necessary information about pedestrian motion and interaction from rich visual information as relevant contextual information for prediction.Second,the decoder is used to decode the input trajectory and contextual information.Finally,to improve the prediction and training capabilities,sampling of the model is studied,using a Quasi-Monte Carlo method as a sampling method and improving a loss function that helps the model converge.The proposed model achieved good results with an average ADE of 0.23 and an average FDE of 0.37 on the public ETH and UCY datasets.(2)In order to improve the efficiency of visual feature extraction and further improve the accuracy of the model’s prediction of pedestrian future trajectories,this thesis studies visual feature extraction in the field of pedestrian trajectory prediction,proposes a Dual Coordinated Attention and a Conditional Dual Coordinated Attention,and constructs a pedestrian trajectory model based on trajectory-constrained visual attention based on the Conditional Dual Coordinated Attention.The model can focus on high information tasks in visual information,use the pedestrian’s historical trajectory as a constraint,and make the network focus on information that is more relevant to the prediction task.Through experiments on the public ETH and UCY datasets,the model further improved the average ADE and average FDE to 0.20 and 0.32,respectively.