| Trajectory prediction technology means that the model predicts the target’s trajectory in the future based on the historical motion state of vehicles,robots,and pedestrians.This technology is also an essential and challenging task in many applications(e.g.,robot navigation,autonomous driving,and intelligent surveillance systems).The key to the safe operation of uncrewed vehicles and intelligent motion robots in complex scenes is to analyze nearby pedestrians and vehicles’ motion intentions and plan a safe and reasonable future movement trajectory to avoid traffic accidents.Since pedestrians are not entities that are only affected by Newton’s laws,the future trajectories of pedestrians cannot be modeled with simple kinematic formulas.Hence,people begin to use deep learning methods to study trajectory prediction problems.However,the existing methods have encountered many challenges,which can be divided into the following two issues:the first one is the redundant problem of modeling motion interaction.Most of the existing methods use pairwise attention mechanisms to model the interaction.When the number is too large,the important distinction between pedestrians becomes smaller,and the interaction information between individuals and groups cannot be captured in different ways.Second,the modeling of target intent is insufficient.Existing methods ignore the dynamic change process of pedestrian intent,resulting in the observation phase not incorporating the pedestrian’s dynamic intent to encode,and the pedestrian’s intent information is not fully utilized.This paper focuses on the trajectory prediction method integrating Spatio-temporal interaction to solve the above problems.It explores four aspects:pedestrian motion interaction,temporal interaction fusion,spatial scene interaction,and target intent interaction.First,in order to simulate the spatial interaction between pedestrians and solve the modeling redundancy problem when there are too many pedestrians in large scenes,this paper proposes a multi-precision interaction model,which includes a novel multi-region interaction module to capture Global interaction,and an additional local interaction module to simulate pedestrian interaction in the local area,and a time-domain interaction fusion method based on attention mechanism is proposed to integrate the time-domain interaction information.The experimental results show that on the ETH-UCY public data set,the results obtained by the method in this paper reduce the ADE(Average Displacement Error)error by 3.8%and the FDE(Final Displacement Error)error by 13.2%compared with the STAR model.Secondly,in order to make full use of pedestrians’ intention information,this paper proposes a multi-dimensional pedestrian trajectory data expression method,which replaces the traditional two-dimensional coordinate representation method.This method can help pedestrians understand the regional information and better mine the coarse precision information of pedestrian movement.Finally,in order to model the pedestrian intention interaction and considering the influence of the spatial scene interaction,this paper proposes a pedestrian intention dynamic analysis sub-network to fit the complete trajectory by combining the pedestrian’s intention information and considering the influence of scene interaction.This sub-network uses multi-dimensional data input to introduce three sub-task loss functions,identifies important scene areas,and dynamically combines scene information.They are using the attention mechanism to fuse the hidden states of the sub-network and the main network to assist the network in updating synchronously.The experimental results show that on the ETH-UCY public data set,the results obtained by the method in this paper reduce the ADE error by 9.6%and the FDE error by 12.2%compared with the Trajectron++ model.In addition,this paper uses the recall rate to evaluate the regional scoring module.The recall rate is 68.6%when recalling Top 1 samples and 99.6%when recalling Top 6 samples. |
PDF Full Text Request