Research On Multi-target Vehicle Detection And Tracking Algorithm

With the rapid development of society and technology,more and more vehicles on the road,intelligent transportation and driverless cars will become the future trend of transportation.In intelligent transportation and driverless car technology,multitarget vehicle tracking has become an indispensable part.Many scholars in China and abroad have carried out key research in the field of multi-objective vehicle tracking.Although achievements have been made,there are still some problems.For example,when more vehicles are tracked,the vehicle is deformed,the vehicles are blocked or the driving environment is complex,the vehicle tracking effect will be worse.In this thesis,around the above-mentioned problems of multi-objective tracking algorithm,the algorithm of vehicle detection and vehicle tracking is studied,and a detection-based multi-objective tracking algorithm is proposed in terms of reducing the rate of missed detection and improving tracking accuracy,etc.The main research work conducted is as follows:The current multi-objective vehicle detection algorithm has high missed detection rate and low accuracy in real-time detection.It is difficult to detect small target vehicles in actual scenes.This thesis optimizes and improves YOLOv5 for these problems.Firstly,in order to reduce the cost of training and reasoning,an improved spatial pyramid pool layer is added to the feature fusion network.The experimental results show that the parameters of the model are significantly reduced after adding the module,which realizes more abundant gradient combination and reduces the cost of training and reasoning.Secondly,in order to improve the accuracy of vehicle target detection,the activation function in the convolution module of YOLOv5 backbone network is replaced by a boundaryless and smoother Mish activation function.The experimental results show that the detection accuracy of YOLOv5 algorithm using Mish activation function is improved.Finally,based on the three-scale prediction of the original network,a scale is added,and a four-scale prediction network is designed to output richer semantic information,so as to better detect multiple targets.The experimental results show that the improved YOLOv5 algorithm has the same detection speed as before,but the accuracy and recall rate are improved,and the detection ability of multi-objective vehicles is improved.There are many vehicles on the actual traffic road,and the adjacent vehicles cannot avoid mutual occlusion in driving,and the speed of the vehicle is not constant.This brings great difficulty to the accurate tracking of vehicles.In view of this,an improved DeepSORT algorithm is proposed in this thesis to solve the above problem.The Kalman filter used in DeepSORT algorithm is used to solve the linear system,and the actual life is a nonlinear system.Therefore,in the tracking phase,the extended Kalman filter is used to predict and update the state of the vehicle target.Then,the deep apparent description network in DeepSORT is retrained to make the features learned by the network suitable for vehicle tracking.Finally,because IOU can not measure the adjacent degree and intersection between the detection vehicle and the tracking trajectory,GIOU is used instead of IOU to improve the matching ability of DeepSORT tracking algorithm.The improved algorithm has improved detection accuracy and precision through comparison experiments,and the tracking failure caused by occlusion is greatly reduced.According to the experimental results,the multi-target vehicle tracking algorithm can adapt to the actual traffic environment and maintain good tracking effect.