Research On Statistical Technology Of Pedestrian Flow Based On Computer Vision

In the application of computer vision,pedestrian flow counting has been a research hotspots,which could obtain multiple pedestrian related information,including the basic information of human flow,pedestrian trajectory and pedestrian retention time within the monitoring area.Pedestrian flow counting not only provides strong safeguard for the management of public place,especially public safety,but also provides reference of performance evaluation for scenic spots,shopping centers and other commercial establishments.However,there are still some problems in pedestrian flow counting at present: higher dimension of pedestrian feature and low detection efficiency of sliding window method,besides,in the course of pedestrian tracking,error tracking will occur when pedestrian cross.Thus the accuracy of pedestrian flow counting is affected.Furthermore,arbitrariness of pedestrian behavior,complex background and the demand for real-time of video processing make the fast and accurate pedestrian flow counting a research difficulty.Therefore,it is of great theoretical significance and application value to carry out the research of human flow statistical technology based on computer vision.Based on the analysis of the advantages and disadvantages of pedestrian detection and pedestrian tracking,this thesis focuses on the study of the framework for pedestrian flow counting,pedestrian feature descriptor and data association of multi-pedestrian tracking.The main contributions of this thesis are summarized as follows:1.The related concepts of computer vision and pedestrian flow counting are introduced.Then the application scope and significance of pedestrian flow counting based on computer vision are introduced.Current situation of the research on human flow statistics are summarized from the aspect of theoretical research and practical application.The advantages and disadvantages of classical algorithms of pedestrian detection,pedestrian tracking and pedestrian counting,three main parts of pedestrian flow counting based on computer vision are summarized,respectively.2.For the inefficiency of pedestrian detection with sliding window method,foreground segmentation and proposal method are researched.First,the principle of representative foreground segmentation methods based on motion analysis are researched.The foreground is obtained with frame difference method,optical flow method and background substraction method,respectively.And the experimental results are analyzed.Then,the principle of representative proposal methods are researched.The proposal windows are generated with selective search method,BING method and edge boxes method,respectively.Also,the experimental results are analyzed.At last,the performance of the three proposal methods are compared.3.Based on the fact that head is the symbol of pedestrian detection under the aerial view,a pedestrian detection method based on continuity of edge and compactness of contour profile is proposed.First,a sample library with whole head and without head is made,then continuity of edge and compactness of contour profile of all samples are extracted to form a two dimensional feature.A pedestrian head classifier is trained with SVM.Then,proposal windows are generated with BING method.The edge responses within proposal window is found with the structured edge detector,continuity of edge and compactness of contour profile are extracted to form a two dimensional feature.At last,pedestrian are detected with the pedestrian head classifier.The experimental results show that the proposed method could detect pedestrian effectively,and the performance of the proposed method is greater than the method based on region symmetry.4.For the problem of higher dimension of traditional HOG feature and higher computational complexity,also,error tracking will occur when the pedestrian cross in the course of pedestrian tracking.A pedestrian flow counting method based on modified HOG feature and SVM is proposed.First,analyzing the space distribution of pedestrian head profile in the detection window,HOG feature of overlook pedestrian head is proposed by optimizing parameter and modifying structure.A pedestrian head classifier is trained with HOG feature of overlook pedestrian head.Then,proposal windows are generated by BING method,and pedestrian within the proposal windows are detected with the trained classifier.The pedestrian are tracked with Camshift algorithm,and pedestrian trajectories are established with the data association algorithm based on Euclidean distance and Bhattacharyya coefficient.At last,pedestrian counting is achieved by analyzing the trajectories crossing the counting line or not.The experimental results show that the proposed method reduces the computational complexity of feature extraction,computation time and the number of error tracking when the pedestrian trajectories cross in the course of pedestrian tracking.The performance of the proposed pedestrian flow counting method is improved.Finally,the research work of this thesis is concluded and further research work is proposed.