Research On Multi-object Vehicle Detection And Tracking Algorithm Based On Deep Learning

With the advent of information technology and intelligence era,intelligent monitoring system has been developed rapidly,through intelligent monitoring system to detect and track vehicles on the road,can effectively realize the process of emergency warning beforehand,timely processing during the incident,accurate evidence afterwards,etc.It can also provide support for vehicle driving status detection and save manpower expenses.Traditional detection and tracking methods are time-consuming and have poor real-time performance;while deep learning methods require high equipment performance and cannot be well applied to lightweight equipment such as cameras,so the algorithm needs to be lightened.This thesis will focus on reducing the number of parameters,computation and model size of the network,reducing the performance requirements of the algorithm on the device and improving the tracking and detection speed of the algorithm,while sacrificing some of the accuracy.The main research contents of this thesis are as follows:(1)Research on algorithms related to object detection and multi-object tracking.The principles of the Faster R-CNN network and Yolov5 s network,the SORT algorithm and Deep Sort algorithm,which are widely used in multi-object tracking algorithms,are studied,and the advantages and disadvantages of these algorithms are compared and analyzed through the results,and the object detection algorithm and multi-object tracking algorithm suitable for this thesis are selected.The data set annotations are formatted and converted for subsequent experiments.(2)Research on improving the vehicle detection algorithm for Yolov5 s network.To address the problem that the current deep learning network cannot better meet the application of lightweight devices such as cameras,the lightweight Mobile Net V2 network is introduced to improve the backbone network of the Yolov5 s object detection algorithm,thereby reducing the model weight size and number of parameters of the object detection network.Considering that more features are lost after the lightweight processing,the CABM attention mechanism is incorporated into the neck network to retain more feature information;in order to improve the detection accuracy based on the lightweight Yolov5 s network,the SPP structural branch is added to increase the perceptual field,and the detection layer is added to the lightweight Yolov5 s network to improve the detection capability for small objects;for the improved The experiments are designed to validate the improved Yolov5 s network algorithm,and stability experiments are designed to verify the stability of the improved object detection algorithm.(3)Research on improving Deep Sort multi-object vehicle tracking algorithm.To address the problem that the feature extraction network model in the Deep Sort algorithm is not very suitable for vehicle appearance feature extraction,it is retrained on the vehicle re-identification dataset and improved with the Shuffle Net V2 network;to address the problem that the Deep Sort algorithm uses a two-stage object detection network resulting in a slow tracking speed,the improved object detection network is combined with the Deep Sort The improved object detection network is combined with the Deep Sort algorithm,and GIo U matching is used to replace Io U matching,which is used to better measure the matching between the object detection frame and the prediction frame;experiments are designed for the improved algorithm,and the performance of the algorithm is compared on the test dataset to verify the running speed of the improved algorithm.Through lightweight improvement processing,the parameters and model weights of the detection algorithm and tracking algorithm in this thesis have been reduced,and the detection speed and tracking speed have been improved to varying degrees.