Research On Key Techniques Of Real-time People Counting Based On Video

In recent years, computer technologies have developed rapidly, and with thedecline in the cost of hardware as well as a variety of video image processingalgorithms be proposed constantly, intelligent video analysis technology in thecommercial, industrial and civilian aspects has reached the practical stage. Theproblem of people counting based on videos is one of the hot spots in the field.Statistics from the flow of people has a high value in life. For example, for exhibitions,sporting events and other gatherings have intensive flow of people, monitoring crowdand controlling the number of participants, can reduce the occurrence of the stampede.Patient flow statistics for hospital outpatient departments in favor of patient flowsummary statistics based on the time that the patient’s treatment, scientific andrational deployment of doctors and nurses on duty time, reducing the waiting time forpatients and so on.This article is the prospect of the segmentation techniques of moving human,machine learning methods to identify the head, tracking and counting the movingtargets, with the hope that getting the statistics of the number of pedestrian at eachchannel or entrance in the surveillance video.First of all, using improved Gaussian mixture model to extract moving objectsfrom the video frame, and then use the external rectangle of the moving targets as thedetection area of the follow-up head. However, due to the use of Gaussian mixturemodel to extract moving targets included the prospect of interference shadows,shadows need to be removed. Compared to deal with the whole image area as a headdetection and counting, this method reduces the workload of the head detection, andcan effectively exclude pseudo-objective in static scene contains similar head,reducing the rate of false detections.Then, analyzing the environment used by people counting system, detectinghuman head region with setting up the camera vertically proposed by Rossi and others.The benefit is that when the pedestrian get close, mutual contacts between the limbswhen occlusion occurs, pedestrian header information is still able to be extractedcompletely. Pedestrian head detection is based on the extraction of Hog feature, linearsupport vector machine classifier as a detection method, is a better overallperformance. Since the classifier accuracy does not depend on the number of support vectors, therefore, using incremental learning training methods and just leaving theclassification of samples which is difficult, can guarantee detection effect whilereduce the number of training samples in final classifier.Finally, tracking and counting the pedestrian in the video scene. Proposed atracking algorithm of moving targets which combine the nearest neighbor matchingand Kalman filter forecasting and detection response model to the head, whichfiltering out some of the dynamic pseudo-target, reducing the miss of head region in asingle frame of video, and the impact on the final statistical result when presenting thefalse detection. When the count, using tripwire-counting method, setting up a virtualline in the scene, as the head inspection box with virtual lines intersect, record thedetection box on the upper left corner of the virtual line distance, by analyzing thesequence distance, accurately determine the direction of pedestrian access anddemographics.This paper presents the prospect of moving target segmentation, the headdetection based on machine learning and the methods of target tracking, all of thesehave achieved specific application. Experiments show that, the statistic of peoplecounting in a period when the pedestrian appears in the video has good practical effect,and can detect pedestrian head effectively.