Trajectory Prediction Based On Multi Scale Information Fusion And GAT

Under the city road,the self-driving vehicle often meets complex dynamic environment.To ensure the rationality and security of decision-making,it is very important to get the estimation of the behavior and future trajectory of other participants in the environment.However,the congested traffic environment and complex road structure in urban roads bring challenges to trajectory prediction.This paper focuses on the vehicle,an important road traffic participant,and studies its trajectory prediction method in complex scenes.For each vehicle,the observed historical trajectory of the past is one of the key information to predict its future behavior.The input of a long enough histo ry trajectory can help the model understand the overall trend of the trajectory and the driver’s individual style.Trajectory prediction is a complex dynamic process.Fine movement adjustment is also important to understand the driver’s behavior.Traditional LSTM encoding methods still have some room for improvement in improving long-term memory ability and capturing subtle changes.In addition,the vehicle is not completely free on the road.The structure and change trend of the road ahead directly affect the future track of the vehicle,but the calculation cost of road information extraction based on grid graph is relatively high.The complex dynamic environment of urban road also brings challenges to interactive modeling.In view of the above problems,th is paper proposes a trajectory prediction method based on multi-scale information fusion and graph attention network(GAT).The main research contents are as follows:(1)A multi-scale trajectory information encoding model MS-LSTM is built to fuse coarse-grained and fine-grained sequence information,so as to give consideration to long-term memory ability and capture subtle changes.MS-LSTM is composed of two groups of LSTMs dealing with different time scales.Coarse-grained LSTM is responsible for extracting the overall trajectory information,while fine-grained LSTM is responsible for capturing the subtle changes of motion,and then the two are fused.(2)In the aspect of road information coding,firstly,the road is transformed into a topological graph structure.Compared with raster representation,the storage cost and calculation cost of topological graph are much smaller.In this paper,the Lane Conv operator proposed by Lane GCN for extracting the road information of topological graph structure is improved,and the relative distance information is introduced in the process of multi-scale fusion to improve the effect of multi-scale road information coding.This simple improvement avoids the mutual interference of information in the process of multi-scale information aggregation using extended convolution,and the introduction of relative distance information is conducive to the understanding of road spatial information.(3)Based on gat,the interaction is modeled,and the overall framework of trajectory prediction model is built with reference to Lane GCN.The training and testing model of Argoverse data set is also established.The method proposed in this paper takes the highest scoring trajectory among the multiple tracks predicted each time.The mean end point error of the predicted track and the real track is 0.13 m lower than that of Lane GCN in the 3s prediction cycle,and the trajectory loss rate which indicates the prediction reliability is reduced by 1.55% and 0.94% respectively when top-1 and top-6 are taken,and all other parameters are reduced,The overall performance of the model has been improved.The qualitative analysis of each improved method was carried out,and the ablation experiment was designed to analyze the improvement quantitatively.