Multi-objective Tracking Algorithm Study Based On Deep Learning

In recent years,with the development and maturity of science and technology,especially deep learning technology and computer level,target tracking technology has increasingly become a research hotspot in the field of computer vision,and also has a wide range of applications in various fields such as real life,industrial applications and military.However,in the practical application of multi-target tracking technology,there are often different scales of targets,background interference and occlusion in the video images,and these problems put forward high performance requirements for the tracking algorithm.Aiming at the technical difficulties in the field of multi-target tracking,considering the real-time nature of the overall tracking task and the requirements of tracking accuracy,this paper determines the multi-target tracking strategy based on detection,and uses the combination of one-stage target detection algorithm Yolov5 and DeepSORT tracking algorithm as the basic research content of the overall multi-target tracking algorithm.In the detection stage,in order to improve the detection effect for targets of different scales in the image,the feature pyramid structure of the Yolov5 target detection algorithm is replaced by the AF-FPN structure;to address the problem of interference of the image background with the target information,the CBAM attention mechanism is added to the detection head structure to improve the ability of the network model to focus on the target features;to reduce missed detection and multiple objects in one frame when the target is occluded,the WBF weighted fusion frame method is used to improve the NMS non-extreme suppression.In the tracking stage,in order to address the problems of shallow layers of the feature extraction network and simple residual structure in the DeepSORT algorithm,the optimization of the residual structure and the improvement of the network structure are proposed to improve the feature extraction ability of the target and further improve the tracking matching accuracy.The experimental data and algorithm results on VisDrone dataset show that in the target detection stage,compared with the original algorithm,the improved ACW-Yolov5 algorithm in this paper improves the overall mAP＿0.5 by 6.274%,mAP＿0.5:0.95 by3.09%,precision by 1.182%,and Recall by 4.919%.In addition,the map＿0.5 of the more important vehicle and pedestrian targets are improved by 8.3% and 9.6%respectively.In the target tracking stage,the improved DeepSORT tracking algorithm in this paper improves MOTA by 8.15% and MOTP by 9.78%,and decreased ID switch by24.61% compared with the original algorithm.The matching effect is improved,the ID switch is reduced,and the overall tracking effect is effectively improved.