Research On Traffic Scene Multi-target Tracking Methods For Unmanned Systems

Recently,unmanned system has become a part of people’s life,especially in the field of intelligent transportation.Two kinds of unmanned systems,unmanned vehicle and unmanned aircraft,have gradually become traffic participators and traffic regulators.However,the traffic scene is complex and changeable,which brings great challenges to the unmanned system based on vision.Therefore,based on the in-depth study of multi-target tracking methods in traffic scene,this dissertation proposes multi-target tracking methods in traffic scene based on target detection,and realizes the multi-target tracking methods based on the perspective of unmanned vehicle and unmanned aircraft,which have practical significance for improving the multi-target tracking ability of unmanned system in traffic scene.The specific research contents are as follows.1.The specific research contents and innovations for unmanned vehicles are as follows:(1)Aiming at the problem that traffic targets vary greatly in scale and the actual movement is complex,so it is difficult to describe the movement shape of the targets by using a single model.A multi-target tracking algorithm based on the YOLOv3 and the interactive multi-model is proposed.Firstly,the algorithm uses the improved YOLOv3 model for multi-scale target detection,and the result of target detection is used as the input of subsequent tracking.Secondly,the interactive multi-model is utilized to calculate the motion state information of the target at some points,and then the target matching is carried out according to the combination optimization algorithm.Finally,the matching object detection boxes are found for each predicted target box.The experimental results show that the MOTP index of the algorithm on MOT16 dataset is 3% higher than that of the offline methods.(2)Due to the influence of the anchor mechanism,it is difficult for the detectors to estimate the accurate boundary,which leads to the poor performance of multi-target tracking.Therefore,the multi-target tracking algorithm without the anchor 3D target detection is proposed.The algorithm adopts no anchor mechanism,which is not affected by the non maximum suppression when the detection frames are merged.The detection model is applied to detect the image and the previous frame,locate the target and predict their association with the previous frame.Experiments show that the multi-target tracking method is effective in 3D multi-target tracking.(3)Aiming at the problem that the targets cannot be continuously tracked when RGB image is used to deal with the occlusion of the targets and the frequent change of the targets in the scene,the multi-objective detection and tracking algorithm based on the multi-modal fusion and the graph neural network is proposed.The algorithm uses the improved sparse data hybrid enhancement method to fuse RGB image and point cloud data,and uses multi-mode 3D target detection method to obtain the target area,and then uses graph convolution network to track multiple targets.The experimental results show that the multi-objective tracking method has achieved the same effect as the most advanced multi-objective tracking method,and has a good performance in MT and IDS.2.The specific research contents and innovations for UAV are as follows:Aiming at the problem that the number of vehicles in the traffic scene is unknown and changes with time,and the low power demand of UAV application,a lightweight multi-target tracking algorithm is proposed based on Poisson multi Bernoulli hybrid filter.Firstly,an improved lightweight multi-target detection algorithm is adopted;Secondly,multi-target tracking is carried out based on Poisson Multi-Bernoulli Mixture filter.The experimental results show that the algorithm can not only effectively suppress the frequent ID transformation problem caused by the repeated occurrence of targets,but also obtain good multi-target tracking accuracy and multi-target tracking precision.Moreover,the algorithm has only 6M parameters,and has advantages in portability and low power consumption.This dissertation mainly focuses on two types of unmanned systems: unmanned vehicle and unmanned aircraft,and proposes corresponding tracking methods for multi-target traffic scenes from different system perspectives.The experimental results show that the proposed algorithm has the ability to track multiple traffic targets from the perspective of unmanned vehicle and unmanned aircraft,and can provide theoretical and technical support for the multi-target tracking technology of traffic scene for unmanned system.