Research On Cross-camera Personnel Tracking Algorithm

Nowadays,with the rapid development of social economy,monitoring devices can be found everywhere in people’s daily life.On campus,in small indoor environments such as classrooms and conference rooms,people are prone to occlusion when walking,which brings great difficulties to the realization of cross-camera tracking.This thesis researches a cross-camera person tracking algorithm,which mainly involves three aspects,namely personnel target detection,personnel target tracking and personnel reidentification.Focusing on the problem of low camera tracking accuracy caused by occlusion,this thesis improves the personnel target detection algorithm and the personnel target tracking algorithm.The specific work contents are as follows:In terms of personnel detection,this thesis mainly improves the YOLOv5s model,the smallest model of YOLOv5,for the problem of low detection accuracy caused by occlusion.First of all,the attention mechanism SE module is integrated into the Backbone of the YOLOv5s model,thereby enhancing the feature extraction ability;Secondly,the CIOU loss function is used to replace the original GIOU loss function,which speeds up the regression rate of the bounding box and improves the positioning accuracy.According to the experimental results,it is proved that the loss function is changed from the original GIOU to CIOU,the accuracy and recall rate are increased by 0.13%and 0.46%respectively,and the mAP value is also increased by 0.14%;After YOLOv5s adds the attention mechanism and improves the loss function,the accuracy is increased by 6.6%.In terms of personnel target tracking,this thesis will use the DeepSORT multi-target tracking algorithm to track the personnel target under the camera.Aiming at the problem that DeepSORT’s ReID module has insufficient ability to extract human features due to personnel occlusion,this thesis replaces DeepSORT’s representation extraction model with the ReID model trained by FastReID.According to the experimental results,it is proved that when occlusion occurs among people,the improved YOLOv5s combined with the improved DeepSORT effectively reduces the number of identity changes(IDs)caused by occlusion during the target tracking process.Regarding personnel re-identification,this thesis uses the FastReID algorithm to re-identify the target personnel after crossing the camera range.Compared with other personnel reidentification methods,FastReID has a high recognition accuracy,which can meet the accurate re-identification of people across cameras.After experimental verification,the cross-camera personnel tracking algorithm researched in this thesis can achieve cross-camera tracking of a specific personnel.In addition,it can meet the requirements of real-time cross-camera tracking in a hardware environment with strong computing power.