Key Target Person Detection And Tracking Based On Fragmented Video Information

Key target character detection and tracking in real-world complex scenes and fragmented information acquisition situations put higher demands on the speed and accuracy of the algorithms.Currently,one-stage target detection algorithms such as YOLO and multi-target tracking techniques such as Deep Sort have the advantages of speed and accuracy,but the detection and tracking of key target characters in special scenes are affected by concealment,occlusion and few samples,and face problems such as target misdetection,target miss detection and target loss.This dissertation focuses on solving the above problems by optimizing YOLOv5 and Deep Sort to achieve efficient detection and tracking of key target characters in special scenes.The dissertation firstly studies the key target person detection algorithm based on YOLOv5 to improve the detection of hidden and obscured key target persons.Then explore Deep Sort-based key target person tracking algorithm,combined with YOLOv5-based key target person detection algorithm to achieve accurate tracking of hidden and obscured people,and finally design and implement a key target person detection and tracking system based on fragmented video information.The innovation of the work in the paper is mainly reflected in the following three aspects:(1)A key target person detection algorithm based on YOLOv5 is designed.The algorithm improves YOLOv5 in three aspects: attention mechanism,loss function,and feature pyramid network.To address the problem of losing effective features of key target person under hidden and occluded situations CBAM attention mechanism is introduced and added behind the three effective feature layers 256,512 and 1024 of the backbone network output to achieve effective extraction of key features;to solve the problem of small number of samples obtained in special scenes,the original BCE Loss is replaced by Focal Loss to solve the The problem of low classification accuracy of categories with few samples;To achieve effective fusion of key features,the weighted bi-directional feature pyramid network(BiFPN)is introduced to replace the original FPN+PANet to fully utilize the effective features.The model is retrained to detect potential key target characters in images,and the experimental results on the Crowd Human dataset show that the improved YOLOv5 algorithm is more effective in detecting characters in complex scenes such as occlusion,hiding and overlapping,which are often key characters.(2)A Deep Sort-based algorithm for key target character tracking is designed.The algorithm improves Deep Sort from two aspects: IOU matching and deep feature extraction network.In feature extraction,the original deep feature extraction network is replaced by the improved Res Net50 network to achieve effective feature extraction of key target characters that are occluded and hidden;after cascade matching,the original IOU matching algorithm is replaced,and CIOU is used to solve the problem of poor effect of IOU matching algorithm in occluded and overlapping scenes,and to improve the matching of detection frame and prediction frame accuracy.The key target person detection algorithm based on YOLOv5 and the key target person tracking algorithm based on Deep Sort are fused,and the optimized YOLOv5 algorithm is used as the detector and the optimized Deep Sort algorithm is used as the tracker,and after retraining the deep appearance extraction model on the Market-1501 dataset,the key target The robustness of the character tracking algorithm is verified on the MOT-16 dataset.Even the hidden or obscured crowd can still be tracked and the ID is not easy to change.(3)The key target person detection and tracking system based on fragmented video information is designed and implemented,and the above detection and tracking algorithms are deployed on the cloud server,and the target detection and target tracking functions can be executed separately on the client side,and the target detection results can be obtained by uploading the images to be detected to the server side,and the target tracking results can be obtained by uploading the videos to be detected to the server side,and the results can be displayed on the visual front-end page.The test results show that the improved YOLOv5 detection algorithm and the improved Deep Sort tracking algorithm can solve the problems of target miss detection and target tracking loss in complex scenes,and the function and performance can better meet the requirements of practical applications.