Pedestrian Head And Shoulders Detection Method Based On Improved ViBe And Machine Learning

Nowadays,video surveillance applies widely to a lot of public occasions such as banks,supermarkets,stations and schools.However,the existing video surveillance system is only in the video recording stage and can't capture the clear appearance of the moving pedestrians.This brings some difficulties to the criminal investigation by policemen.Therefore,it's of great significance to study pedestrian detection and clear appearance capturing of pedestrians in video surveillance.Head and shoulder detection is the key step for capturing a clear appearance of pedestrians.It aims at obtaining head and shoulder positions of the pedestrians accurately and providing reliable preconditions for capturing a clear appearance of pedestrians.This paper applied image processing and machine learning technology to conduct on the study of detecting head and shoulders in public occasions.The main research content was divided into three modules which were moving target detection module,pedestrian detection module and master-slave camera linkage calibration module.The specific work and innovation of this paper are as follows:(1)For the requirement of the clear capturing for pedestrians,this paper used fisheye camera and PTZ camera to design the master-slave monitoring system.For the linkage requirement between master and slave cameras,this paper applied spatial calibration method based on digital fitting by generating the look-up table based on the pixel position of the sample point in the fisheye camera and the rotation angle of the PTZ camera.Finally it completed the linkage calibration and achieved the purpose of the linkage between the master and slave cameras.Verification showed that this can conduct a linkage every 5 milliseconds after the pedestrian head-shoulder detection and clear appearance capture system was completed.(2)For the problem of "ghost area" and moving target shadow interference when detecting the moving target by the traditional ViBe algorithm,this paper proposed to combine perceptual hash algorithm with ViBe algorithm for the suppression of " ghost area " and use “Gaussian Laplacian Difference Method” based on image's RGB color information to eliminate moving object shadow.Compared with the traditional ViBe algorithm which completed the suppression of "ghost area " in the 1315 th frame of video,the improved Vi Be algorithm only completed the suppression of the "ghost area" in the 15 th frame of the video,and there is no interference of moving target shadow.(3)For the problem of high false detection rate in pedestrian detection process,this paper proposed a two-stage pedestrian head and shoulder detection algorithm.In the first stage,this paper used the cascade classification algorithm based on AdaBoost thought training HOG feature to generate the “First head-shoulder detector”,detecting the "candidate area" of pedestrian shoulder-shoulder parts.In the second stage,this paper used the SVM classification algorithm training the ORB feature to generate the "Second head-shoulder detector ",detecting the "candidate area" for the second time and taking the result of the second stage as the final result.Experiments showed that the accuracy of two-stage detection algorithm was 80.86% and was nearly 10% better than the traditional HOG+AdaBoost detection algorithm.