Research On Visual Human Object Detection And Tracking Technology Based On Deep Learning

Target detection first filters the target class from the video or image,and then the target detection algorithm marks the specific position of the target in the video or image.Finally,the class name is labeled by the classifier.The YOLOv3 target algorithm has received wide attention because of its fast detection speed and high accuracy,but under the factors of complex scenes,target scales and pose diversity,the algorithm may miss,false detected or repeat detection for some small targets.To avoid the above situations,the YOLOv3 algorithm is improved,and the improved algorithm is also combined with the Deep SORT algorithm to enhance the tracking effect of the joint algorithm.The main research contents of this paper are as follows:(1)Study the target detection problem.To address the problems of low detection rate of YOLOv3 for small targets,failure to respond to the intersection of prediction frame and target frame when the Io U values are the same,and frequent false suppression of traditional NMS algorithm for occlusion cases,the residual block and neural network structure of YOLOv3 are firstly improved.Secondly,the SPP module is introduced and the improved residual blocks are fused into the CSPNet module.Finally,DIo U is used as the loss function to replace the Io U loss function,and DIo U-NMS is used as the classifier to improve the detection of small targets and occluded targets in the improved network.(2)Based on the improved YOLOv3 algorithm,the Deep SORT algorithm is combined with the improved algorithm.The joint algorithm first uses the improved YOLOv3 algorithm to detect the target,and then implements the target tracking by Deep SORT algorithm.Deep SORT is based on the SORT algorithm,which uses the wide Res Net network to train the human epigenetic features offline and adds the epigenetic features of the tracked target to the matching algorithm such as the Marxian distance,with the purpose of reducing the frequent ID Switch cases and improve the tracking effect on obscured targets.(3)Based on the improved joint algorithm of YOLOv3 and Deep SORT,the study is extended from human target detection and tracking to mask target detection and tracking.First,the person classes in the VOC2012 dataset are extracted using a script,and the collated dataset is used to train and test the improved YOLOv3 algorithm.Second,to facilitate mask wear detection and reduce the workload of prevention personnel,the mask dataset was collected,aggregated and labeled in the experimental phase,and the improved YOLOv3 algorithm was used to test the accuracy of mask detection in complex scenarios.Finally,the joint algorithm is applied to the camera or video,which can basically track and count masked targets and unmasked targets accurately.