Research On Vehicle Trajectory Prediction And Risk Assessment At Intersections

The development of Intelligent Vehicle（Intelligent Vehicle）contributes to the safety,efficiency,energy saving and environmental protection of the urban traffic environment.Therefore,in recent years,intelligent driving technology has received extensive attention and research.However,when intelligent vehicles are driving in a highly dynamic urban intersection environment,they will often make wrong driving decisions and cause serious collisions due to their insufficient knowledge of the possible future movement changes of surrounding traffic agents.In response to this,the main content of the paper is to study the vehicle trajectory prediction algorithm in the intersection environment and build a intelligent car joint collision risk assessment model.The specific research content is as follows:（1）In the second chapter,the paper proposes a traffic light detection and recognition algorithm based on semantic segmentation model（BSSNet+TL2Net）.First,based on the collected 15 types of traffic light dataset and the algorithm experimental verification platform built,the paper compares and analyzes the detection and recognition performance of Faster R-CNN algorithm and YOLOv3 algorithm on the traffic light dataset of the paper.Secondly,considering the accuracy and real-time requirements of traffic light detection and recognition tasks,in the feature extraction of traffic lights,the paper uses the proposed binary semantic segmentation model（BSSNet）to replace the deep convolution model in the traditional detection algorithm.It effectively improves the detection accuracy of traffic light small targets,and avoids the generation of target candidate anchor（Anchor）and nonmaximum suppression processing（NMS）operations,thereby greatly reducing the computational complexity of the model.Finally,through the qualitative evaluation and quantitative analysis of Faster R-CNN,YOLOv3,BSSNet+TL1Net and BSSNet+TL2Net models,the detection and positioning accuracy and real-time performance of the traffic light detection and recognition model proposed in this chapter are verified.（2）In the third chapter,the paper proposes a universal multi-agent future trajectory prediction model Social-WGDAT in the urban intersection environment.First of all,in the encoding layer of the model,in order to effectively simulate the interaction between traffic agents,this paper proposes to use the graph double attention network to extract the characteristic information of the nodes from the agent interaction graph,and calculate the influence correlation between the agents based on the topological attention coefficient degree.Secondly,in the decoding layer of the model,this paper proposes to apply a kinematic constraint unit while decoding the agent’s future motion trajectory through the GRU unit,thereby ensuring the physical feasibility of the agent’s future trajectory prediction.Finally,this paper analyzes the Social-WGDAT prediction model and other baseline prediction models proposed in this paper based on the ADE and FDE trajectory prediction evaluation index experiments,thereby verifying the prediction accuracy of the SocialWGDAT trajectory prediction model.（3）In the fourth chapter,the paper builds an intelligent vehicle joint collision risk assessment model.First,based on the prediction results of the future trajectory of the vehicle in Chapter 3,the paper analyzes the collision probability of the intelligent vehicle within the predicted trajectory range.Within the predicted trajectory,a collision risk assessment model is established based on factors such as the probability of collision,collision time and relative driving speed between vehicles.Secondly,taking into account the uncertain factors outside the predicted trajectory range,build an intelligent vehicle joint collision risk assessment model based on the normal distribution assumption.Finally,an intersection simulation environment was built based on the Carla simulation platform,Carsim vehicle model building software and Road Runner traffic scene building software.And the robustness of the intelligent vehicle joint collision risk assessment model proposed in the paper is verified in unexpected object scenes,sensor failure scenes and different sensing system scenes.