Research On Vehicle Trajectory Prediction In Unsignallized Intersection Scenarios

Accurately predicting the trajectory of surrounding vehicles is the basis for intelligent vehicles to understand the traffic environment and to make independent decisions.There are few existing researches on vehicle trajectory prediction in the scenario of unsignallized intersection,and such studies do not fully extract the temporal hidden features of vehicle trajectory,relying on serial trajectory encoding,and when considering vehicle interaction,the selection of interactive vehicles is unreasonable and the interaction modeling is insufficient,which is only suitable for sparse traffic.When the multi vehicle conflict game behavior in the intersection is frequent and the interaction relationship is strong,the accuracy of the existing methods in this scenario will decline.Taking vehicle trajectory prediction as the theme,this paper proposes an interactive vehicle selection method suitable for unsignallized intersection scenes,designs a vehicle trajectory prediction model based on spatial-temporal features of multi vehicle hybrid interaction,solves the problems of insufficient vehicle trajectory feature extraction and multi vehicle interaction relationship modeling,and proposes a prediction trajectory generation method based on multi optimization objectives,The high-precision vehicle trajectory prediction in the unsignallized intersection scene is realized.Firstly,aiming at the problem of interactive vehicle selection,this paper analyzes the road structure characteristics of representative unsignallized intersections,the conflict points or interleaving points at intersections,and the possible interaction behavior between vehicles,and puts forward an interactive vehicle selection method for intersections and roundabouts.The neighbor vehicles of the predicted target are captured through a variable length neighbor agents capture grid.The interactive area division methods for intersections and roundabouts are proposed to screen the interactive vehicles that are far away but interact with the prediction target.Secondly,aiming at the insufficient problem of vehicle trajectory feature extraction and multi vehicle interaction modeling,a vehicle trajectory prediction model based on spatialtemporal features of multi vehicle hybrid interaction is designed.The vehicle trajectory is encoded in parallel,and the spatio-temporal interaction relationship of multi vehicles is modeled to improve the accuracy of prediction trajectory.According to the temporal dependencies of each trajectory point in the trajectory,two relative positional encoding methods that can reflect this dependencies are designed,which are combined with the multi-heads attention module in the trajectory feature encoder based on Transformer to encode the trajectory in parallel and preserve the temporal dependencies among the trajectory points;An interactive feature encoder is designed,which encodes the trajectory data of interactive vehicles in parallel through the multi-heads attention module with mask mechanism,and captures the hidden features of spatial-temporal interaction;A fusion block including two linear layers and two fully connected layers is designed to fuse the hidden features of the trajectory of the predicted target and the hidden features of the interaction;A prediction trajectory decoder based on Transformer is designed to generate the future trajectory of the predicted target according to the understanding results of the two encoders.In addition,aiming at the problem that the diversity and great difference of road structures affect the prediction accuracy,a prediction trajectory generation method based on multiple optimization objectives is designed,which transforms the road structure constraints,speed constraints and future trajectory constraints（the sequence of trajectory points predicted by the trajectory prediction model）into optimization objectives,and establishes the optimization objective function,Taking the quartic polynomial function as the trajectory basis function.Then,the final predicted trajectory is solved iteratively.Finally,the INTERACTION dataset is used to verify the effectiveness of trajectory prediction method and prediction trajectory generation method.Firstly,the dataset is preprocessed,including data smoothing and filtering,trajectory segment segmentation,static trajectory segment elimination,mark start symbol,etc.,and 131328 samples are obtained（each segment is about 6 seconds）.Then the algorithm code is programmed by Python + Pytorch,and the accelerated training of the model is realized in combination with GPU computing platform.The results show that the longitudinal ADE and FDE are 0.53 m and 0.94 m respectively,and the transverse ADE and FDE are 0.73 m and 1.48 m respectively.