| With autonomous vehicles on the road testing,some prominent problems were exposed.For example,in terms of perception,due to the complex real traffic environment and increasingly complex perception algorithms,it is prone to instability during target tracking,leading to the wrong understanding of the scene information;in the aspect of behavior prediction,the current driving behavior is too conservative and does not have the ability to interact with other cars.According to the above problems,the paper studies the design of vehicle target tracking system,interaction target selection and interaction behavior prediction.(1)Design of the vehicle target tracking system.Select the coordinate system of tracking system,design data association algorithm for tracking,Unscented Kalman Filtering(UKF)tracking algorithm based on constant acceleration(CA)motion model,and management rules for trajectory.The effectiveness of the designed target tracking system is validated through simulation and experiment.The results show that the designed vehicle target tracking system can effectively solve the error detection and missed detection problems of the sensor during the target tracking,and improve the tracking stability of the vehicle.(2)Vehicle interaction target selection.In order to improve the real-time,stability and effectiveness of the work of the autonomous driving system,it can judge whether there is an interaction between the vehicles to filter out the unimportant tracking targets returned by the tracking system.Based on the analysis of target selection behavior of human drivers interacting with vehicles,and based on trajectory classification and risk assessment,an interactive target selection method for autonomous driving vehicles was established for intersection scenarios.For trajectory classification,the effectiveness of Hausdorf(Hausdorff)distance,Time Gauge Algorithm(DTW)and the longest Common Subsequence(LCSS)during traj ectory classification is analyzed.The results show that the LCSS has excellent classification performance with an accuracy of 82.25%.For the collision risk assessment,the driving reference path is established based on the road structure,representing the position and status of the vehicle in the Frenet coordinate system,and taking the relative position relationship and collision time of the vehicle as the basis for judging whether there is an interaction.The established vehicle interaction target selection method is verified on the real traffic flow data,and the result shows that the selected vehicle interaction target is more reasonable.(3)Vehicle interaction behavior prediction.In order to provide autonomous vehicles with similar scene prediction capabilities to human drivers to achieve driving in collaboration with human drivers in interaction scenarios,a deep generative model conditional variational autoencoder(CVAE)is used.The rationality of the modeling using CVAE is analyzed and the prediction model of vehicle interaction behavior is established.The model is trained using vehicle interaction data in real traffic scenarios and converted into the Frenet coordinate system.The test results show that the behavior prediction model has good prediction accuracy and considers the uncertainty of human driving behavior;using the Frenet coordinate system to describe the behavior prediction model has good generalization.Comparing the behavior prediction results for considering and without regard for the interactive target vehicle,and show that the behavior prediction is more accurate and reasonable when the interactive target vehicle is considered. |
