Research On Multi-face Detection And Tracking Algorithm Based On Video

Face detection and tracking technology is widely used in smart store systems,smart retail and video surveillance.After studying some face detection and tracking algorithms,this paper proposes to train the face detection model using YOLOv3 algorithm.In view of the shortcomings of the KCF tracking algorithm in the tracking process,the following improved methods are proposed:(1)A strategy for multi-feature fusion is proposed.The response results calculated by the HOG feature and the color feature response result are fused according to a certain weight.And select the location of the maximum value of the response after the fusion as the location of the tracking target.(2)Propose a new model update strategy.When the tracking target is occluded,if the model is updated with the previous weight,the tracking target will be lost due to the background pollution model.Therefore,this paper proposes that when it is judged that the tracking target is lost,the update of the model can be stopped,and the model is continuously updated until the target is found.(3)Propose a scale update strategy.Since the KCF tracking algorithm can not solve the scale problem of the tracking target,this paper draws on the scale filter of the DSST algorithm to solve the multi-scale problem based on the improved KCF algorithm.Finally,this paper proposes a multi-face detection and tracking algorithm,combining face detection and tracking,and proposes two methods related to detection and tracking.One is to mark the face that is being tracked,and then detect whether there is a face in the marked image.Assign a new tracker trace if it exists.The second is to use the distance accuracy and the intersection ratio to calculate the correlation between the detected face and the tracked face.If the detected face does not have a tracking face associated with it after the calculation,it is determined as a newly detected face and given a new tracker.Then use the sobel edge detection operator to get a clearer face image from the second association scheme to complete the face recognition.