Research On Target Detection And Tracking Algorithm Based On Video Surveillance System

Target detection and tracking is one of the important research directions in the field of computer vision.Its main purpose is to achieve visual perception and effective tracking by determining the target position in the image.However,deep learning methods require a lot of computing power,which makes them difficult to apply to mobile devices.At present,deep neural networks have become larger and more complex,and many of these models are not suitable for mobile porting,which has become more prominent.Based on the research of target detection and tracking algorithms in the surveillance system,this thesis deeply explores the multi-target tracking algorithms in the field of computer vision,and focuses on the detection-based multi-target tracking strategy,including online and real-time deep association metric tracking.A target detection and tracking algorithm model based on machine vision is proposed,which is applied to the detection and tracking tasks in the monitoring system.1)By testing the performance of various mainstream detection algorithms,the YOLO series detectors are selected.In the main network CSParknet53 of YOLOv5 algorithm,there are problems such as large number of model parameters and high hardware equipment requirements.It mainly includes : replacing CSPDarknet53 with a lightweight Shuffle Netv2 as the backbone network,and adding Stem Block and Coordinate Attention to compensate for the loss of accuracy due to model lightweighting.The algorithm performance test is carried out on the self-made data set.The experimental results show that the network parameters of the improved YOLOV5 model are only one tenth of the original,and the m AP value is reduced by 7.01%.Under the premise of controlling the detection accuracy,the detection speed of the model is increased from 74 FPS to 88 FPS.2)The algorithm improves the Sort algorithm,retains the Kalman filter trajectory prediction and Hungarian data association method,and introduces a convolutional network to improve the description ability of deep features,thereby improving the reliability of the model.In addition,this paper also improves the Deep Sort algorithm,using the complete intersection over union(CIOU)instead of the traditional intersection over union(IOU)to improve the accuracy of target matching.Finally,the improved YOLOv5 is used as the detector of the improved Deepsort algorithm model to achieve tracking.The accuracy and accuracy are increased by 2.2 % and 0.9 % respectively,and the speed reaches 64 FPS,which improves the tracking speed and accuracy of the model.