Research On Road Vehicle Tracking Methods Based On YOLOv5 And Deep Sort

Traffic monitoring is an important tool to provide data support in order to control traffic flow,solve traffic congestion and rationalise road planning.Vehicle detection and tracking is an important part of traffic monitoring,and although its research has made great progress,there is still room for improvement in simplifying the network model and reducing costs,and it has become one of the research hotspots of Intelligent Traffic System(ITS).Deep learning based multi-target tracking is an important technology in traffic monitoring.This technology can solve the problem of unstable target acquisition in complex situations by first extracting the features of the vehicles detected in the surveillance video,and then performing accurate and stable tracking.In recent years,most deep learning based detection techniques have been used to improve the feature extraction capability of the network by deepening the depth of the network.However,as the network deepens,it leads to an increase in the number of network model parameters,slows down the computation speed,and also imposes more stringent requirements on the equipment.Therefore,with limited computational resources,there is a need to design a network model with low number of parameters,high speed and guaranteed accuracy for deployment on mobile devices.In this thesis,we propose a YOLOv5-based vehicle detection and tracking algorithm,which is a lightweight network improvement,taking into account the large number of parameters in the existing model and the need to increase the speed on the CPU,respectively.The main process of this algorithm is to first locate and tag targets using detectors,then perform feature extraction on the tagged targets,and finally achieve multitarget tracking through data association matching.The Rep VGG network is introduced to lighten the backbone network of the original YOLOv5 detection model to address the problem of large number of parameters in the existing model;the coordinate attention mechanism is introduced to obtain the target direction and position sensing information,thus overcoming the problem of inaccurate tracking of position information in the spatial direction of the target in the traditional tracking network and improving the feature extraction capability of the network;the improved YOLOv5 network is used as the Deep Sort front-end target detector,using the Kalman filter algorithm for predictive tracking of targets,and using the Hungarian algorithm for cascade matching.Experimental results show that compared to the original YOLOv5-Deep Sort,our proposed method decreases the number of parameters by 78.3MB while maintaining the same detection tracking accuracy,which is 11.4% of the number of parameters of the pre-improvement algorithm.For CPU devices,a CPU lightweight network PP-LCNet is introduced to lighten the detection network,and a convolutional attention mechanism module is introduced to overcome the problem that traditional tracking networks tend to ignore the position information in the spatial direction of the target and improve the tracking accuracy.The experimental results show that compared with the original YOLOv5-Deep Sort,the detection time of our proposed method is 12.5% of the original detection time,the tracking time is 17% of the original tracking time,and the number of network parameters decreases by 79.61 MB,which is only 10% of that before the improvement.In addition,this paper has done comparative experiments with other algorithms on the UA-DETRAC dataset to demonstrate the practicability and superiority of our algorithm.