Research On Trajectory Prediction Method Of Traffic Participants In Complex Urban Traffic Scenarios

With the increasing complexity of urban traffic scenarios and the gradual application of intelligent driving system,the urban traffic safety issues has attracted much attention.Traffic participant trajectory prediction aims to predict the future position of dynamic agents such as pedestrians and vehicles based on the historical observation trajectory.It can provide planning suggestions and decision-making guidance for autonomous vehicles,intelligent monitoring systems,mobile robots,etc.,and improve traffic safety in complex traffic scenarios.It is a key task in the fields of pattern recognition and intelligent transportation.The current trajectory prediction methods have problems such as insufficient social interaction modeling,uncertain future motion trajectories,and poor long-term prediction performance.This paper conducts an in-depth research on the above issues.The main work and innovations are as follows:(1)To address the problem that the existing models cannot fully model the social interaction between agents,this paper proposes an agent trajectory prediction network based on spatial-temporal interaction attention(Spatial-Temporal Interaction Attention based Trajectory Prediction Network,STIA-TPNet).This paper designs a spatial-temporal interaction attention module,which consists of two branches:temporal flow and spatial flow,and explicitly extracts interaction information and spatial-temporal dependencies between agents.Firstly,the temporal interaction attention is used for temporal dimension to analyze the importance of the historical trajectory of different moments to the future movement of the target agent.Secondly,spatial interaction attention is used for spatial dimension to capture the interaction between neighboring agents and target agents.Finally,the spatial-temporal interaction information is fused,so that more important information can be selected adaptively in the time domain and the space domain,which enhances the interpretability of the model to the spatial-temporal interaction process.(2)In order to solve the problems of uncertainty of future trajectory and poor long-term prediction performance due to the inherent characteristics of agent motion,this paper proposes an agent trajectory prediction network based on endpoint driven and reverse enhanced decoding(Endpoint Driven and Reverse Enhanced Decoding based Trajectory Prediction Network,EDRED-TPNet).Firstly,considering the randomness of the agent movement and the dynamics of the number of agents in the scenario,a spatial-temporal scene graph is established to capture the dependency of time and space simultaneously.Then,an endpoint driven module is designed,which learns the final motion intention of the agent before predicting the complete trajectory.This potential endpoint can be associated with the observed historical trajectory and combined with reasoning to obtain the complete motion trajectory,which provides guidance for trajectory prediction.Finally,a reverse enhanced decoder is proposed.By blending the reverse and forward hidden state vectors to enhance the influence of hidden state vectors,the performance of long-term prediction is improved,and the problem of large long-term prediction errors caused by information loss and cumulative errors is effectively improved.(3)Due to the fact that the current public datasets in the field of trajectory prediction are rarely oriented to real complex urban traffic scenarios and social interaction information among various traffic participants cannot be fully considered,and trajectory data for urban traffic scenarios in China is relatively scarce,this paper uses UAV to collect data above intersections and constructs a trajectory prediction dataset(UAV-captured Trajectory Prediction Dataset,UTP).UTP dataset includes pedestrians,cyclists,vehicles and other types of traffic participants,covering complex and diversified urban traffic scenarios and rich agent interactions,which provides a data basis for the research of trajectory prediction task.The proposed method is tested on the constructed UTP dataset and the public INTERACTION dataset respectively,and compared with other methods in terms of average displacement error and final displacement error matrics,proving the effectiveness of the proposed method in this paper.