| The bad behavior of personnel on duty such as sleeping on duty and leaving the post without authorization may cause serious accidents in the key duty room(such as fire control room,railway operation room,intensive care unit,etc.).Therefore,it is very important to monitor the working state of personnel.An intelligent data analysis system is built based on monitoring video to extract the state of monitored human body in real time and accurately.It can not only effectively avoid accidents,but also save a lot of manpower and financial resources and promote the development of social intelligence.Among them,the detection,tracking and state recognition algorithms for human body are the core of intelligent data analysis system.The human body has the characteristics of rigidity and flexibility at the same time with different appearance.And human body clothing,imaging size,complex background and lighting conditions have great influence on the accuracy of human detection and recognition.Therefore,human recognition has always been a difficult point in the field of computer vision.According to the task requirements of duty room personnel working state discrimination,In this paper,the state recognition of human object is defined as mobile recognition,and the research target is the whole motion,micro motion and static state recognition of the human object,and the research of human body detection,tracking and state recognition algorithm is carried out.The main research work is as follows:In the first of all,this paper summarizes the basic knowledge of Faster RCNN human object detection from human data setting preparation,evaluation criteria of human detection algorithm,regional recommended network structure and learning rate setting.And then human region precision inspection has been realized with the method of combining average background modeling with the Faster RCNN detection results to extract the background so as to avoid that the accuracy of subsequent tracking algorithm is reduced because of too much background in the target block.The experimental results show that the human body detection rate of the trained Faster RCNN model can reach 96%,which is superior to the traditional human detection method under the same conditions.Second,the human body tracking algorithm based on multi-strategy fusion is designed for the characteristics of human multi-joints and the diversity of scales and dress colors.The color classification module is designed based on the characteristics of small differences on recognition due to the large difference between the color and background of human clothing,so that samples can be classified by using the statistical features of color,the color distribution and point-pair similarity.The HOG feature is used to design the tracking module and the feature function is updated by using the classification result to avoid the drift of the tracking module effect with time,and the multi-strategy fusion is realized by synthesizing the results of the classification module and the tracking module.Moreover,two groups of laboratory videos are used to verify the algorithm.The results show that the tracking algorithm has long time tracking ability.The experimental results show that the Multi Strategy tracking algorithm designed in this paper has good adaptability to the problems such as occlusion,deformation,size change and so on.The average value of tracking results and targets in the tracking process is about 0.8,and the average center distance error is less than 6 pixels.Finally,the human body detection and tracking results are used to extract the human body region,and count the time movement information of human body and discriminate the human movement state,thus extracting the working state of the human body.According to the local motion characteristics of human body,this paper also proposes a fine motion detection algorithm based on spatio-temporal filtering.In addition,the motion region is extracted by three frames difference rough,the motion feature points are extracted by hog tracking and fb error method,and the motion,color similarity and spatial connectivity are used to cluster the motion.In order to extract the moving pixels accurately,the cluster of pixels in the moving region is realized.The experimental results show that the proposed method can accurately extract the target motion information and achieve efficient human state recognition. |
PDF Full Text Request