Vehicle Detection And Tracking Based On Depth Feature

As an advanced traffic management technology,intelligent transportation system is paid more and more attention by people.With the improvement of computer hardware level,the problem of real-time calculation caused by many complex algorithms has been gradually broken through.The vehicle detection technology studied in this paper is an important part of the road safety monitoring system,which is of great significance for the development of intelligent transportation system.Firstly,the feature extraction method of vehicle is studied.In order to reduce the interference of natural environment(such as the change of light intensity and shooting angle,the diversity of background,etc.)and save memory and speed up the calculation in the process of training,this paper proposes to fuse the three layers of features after CNN and input them into SVM classifier after dimension reduction by PCA Vehicle identification method.This method is based on Alexnet model.By comparing the test accuracy of different parameters through experiments,a convolutional neural network model is constructed to obtain the optimal vehicle recognition system.The serial fusion method is used to fuse the three layers of vehicle feature map after the extracted network,and a vehicle feature vector with sufficient information is obtained through principal component analysis to improve the accuracy and reality of vehicle recognition In order to enhance the generalization ability and error correction ability of the model,CNN's SoftMax is replaced by SVM classifier to realize vehicle recognition.The experimental results show that compared with the traditional feature extraction method,this method significantly improves the speed of vehicle recognition and classification accuracy,and has good robustness.Secondly,the algorithm of target detection and tracking is studied.In order to ensure the real-time performance of vehicle detection system,YOLO is proposed As the basic model of vehicle target detection,V3 uses K-means + + clustering algorithm to cluster the target candidate frame,so as to improve the detection ability of the model for unknown objects,and adds a feature extraction layer in the shallow layer of the network,which can extract more precise vehicle features,improve the detection performance of the network for road vehicles,in addition,in order to strengthen the robustness of the network for different targets far and near,While keeping the originaloutput layer of YOLO V3,a layer of output layer is added;then the improved model of YOLO V3 is trained to achieve vehicle target detection.After that,the Deep-Sort algorithm is introduced to track the target.YOLO V3 belongs to the target detection algorithm,and Deep-Sort belongs to the tracking algorithm.The adoption of these two algorithms in the same system will cause the same target to be repeatedly detected and reduce the detection performance.In order to solve the problem of duplicate detection,the Hungarian algorithm is used to correlate and match the detected data and remove the YOLO The experimental results show that the precision and recall rate of the method are 90.16% and 91.34% respectively,which improves the precision and recall rate while ensuring the detection speed,and has good robustness for the detection of different size targets.This paper focuses on three related algorithms of vehicle feature extraction,detection and tracking.Firstly,convolution neural network is used to extract vehicle features.On this basis,YOLO V3 network is improved,and Deep-Sort algorithm is introduced to further improve the detection performance.It is found that the design and improvement of this paper not only improves the accuracy of vehicle detection,but also improves the detection efficiency to a certain extent.