Researchresearch On Multi-target Detection And Tracking Algorithm Based On Deep Learning

Multi-target detection and tracking in surveillance camera is the main research direction in computer vision.The development of deep learning has made breakthroughs in the research of multi-target detection and tracking,the continuous evolution of hardware resources make the deployment of algorithms in practical applications conveniently.In the actual application of image detection and tracking,the uncontrollable factors such as multiple targets,complex background and motion changes,require the algorithm with higher detection performance.Based on the theory and technology of deep learning algorithm,this paper aims at both the high-precision and rapid detection algorithms of multiple targets and dense targets with different scales in surveillance video,and the lightweight model suitable for deployment.The research work has been carried out as following:(1)A multi-scale high-precision detection algorithm,named i1-YOLO integrated by attention mechanism is proposed to improve the detection accuracy of small and dense targets.The specific strategies are as follows: a multi-scale detection layer is added to improve the feature extraction ability of small targets and dense targets,so that the model has a stronger ability to obtain in-depth information of images.The CBAM attention mechanism is added to emphasis the important features of the image without inhibiting the effect of non-important features,that is,to pay more attention to the small target information without losing the information of the normal and the large targets.By updating the intersection ratio of detection frame and prediction frame to CIOU,the algorithm is more sensitive to the distance information between prediction frame and real frame,so that the regression ability of the model to boundary frame is better.Based on the self-built target detection dataset,a series of comparative experiments are carried out.Compared with the original YOLOV5 S model,the average accuracy of the proposed network model i1-YOLO is improved by 3%,especially for the detection of small targets and dense targets.(2)A lightweight algorithm combining pruning and knowledge distillation,i2-YOLO,is proposed.The algorithm has the advantages of high precision,fast speed and small size.The specific strategy is as follows: firstly,in terms of data enhancement,Mosaic data enhancement method is used to enhance the image data.Then,L1 regularization is used to constrain the coefficients of BN layer to make it sparse.After the sparse training of the model,the coefficient factors that have weak influence on the output results are obtained,and the parameters of the corresponding layer are trimmed.Experimental results show that the detection speed of i2-YOLO network is improved from 34.4FPS to 70.6FPS under the premise that the model accuracy is decreased by 0.67% slightly,which is 205.2% higher than the original model of YOLOV5 S,the model size is only 39.2% of YOLOV5 S.Therefore,i2-YOLO meets the requirements of embedded onboard deployment and serves as the target detector for multi-target detection and tracking algorithms.(3)An improved deep Re ID model based multi-target tracking algorithm i-Deep SORT is proposed.The specific improvement method is as follows: firstly,the input data is enhanced by image preprocessing module.Then adding the vehicle Re ID features to enrich the prior information of vehicle targets.Finally,the deep Re ID network is adopted to improve the feature extraction ability of the target.Compared with Deep SORT algorithm,the experimental results show that the i-Deep SORT significantly improves the performance of Rank1,Rank5,Rank10 and Rank20.(4)The i2-YOLO and i-Deep SORT areintegrated to construct the multi-target detection and tracking algorithm.In order to test the performance of this algorithm,YOLOV5 S and i2-YOLO are integrated with Deep SORT and i-Deep SORT respectively to obtain four different algorithms.According to the test results of 3 video scenes,i2-YOLO and i-Deep SORT integrated algorithm has the best performance in MOTA and ML indicators.Finally,by building the development environment on NVIDIA Jetson AGX Xavier board,i2-YOLO and i-Deep SORT integrated algorithm is transplanted on it and achieves fast and effective multi-target detection and tracking.The dataset in research(1)and(2)mentioned above contains over 40,000 target detection dataset provided by cooperative enterprises.The dataset of research content(3)is a fusion dataset of the pedestrian and vehicle Re ID dataset from the public resources and the cooperative enterprises,with over 95,000 samples.