Research On Foreground Extraction And Pedestrian Detection And Tracking For Video Surveillance

With the development of computer vision,pedestrian detection in the intelligent auxiliary driving,intelligent monitoring,pedestrian analysis and intelligent robot,and other fields have been widely used.However,because of the complexity of the real life background,pedestrian posture diversity and diversification of shooting angle,there is a big challenge for us to let the pedestrian be extracted from the video fast and efficiently.In this way,the pedestrian detection has been a very hot topic in the research area of computer vision.Aiming at the problem of traditional HOG+SVM pedestrian detection,such as low timeliness and possible false detection,this paper starts with the classical algorithm—HOG+SVM,combines the existing technology and practical application scenarios,and proposes some solutions for these problems.The main content of this paper is shown as follow:1.In terms of foreground extraction of moving targets,our experiment found that the traditional pedestrian detection time loss is mainly concentrated in the HOG feature calculation.Therefore,this paper uses the foreground extraction algorithm to reduce the area of the HOG calculation area,thus speeding up the algorithm.This paper studies two classical foreground extraction algorithms—the Gaussian Mixed Model(GMM)and Vibe.However,the foreground always contains the shadow,some noise points and holes inside objects.So this paper introduce removing shadow,an image morphology method and border scaling to solve this problem.Compared with the previous algorithm,it is found that the proposed algorithm extracts less foreground noise and is more accurate.2.This paper summarizes three general feature models for object tracking,the tracking algorithm based on template matching,Pyramid Lucas-Kanada(LK)optical flow tracking algorithm based on feature points and the tracking algorithm based on Kernel ized Correlation Filters.Then we do some experiments to track human face using algorithms above.3.Taking into account the problem of low speed of pedestrian detection with a HOG-SVM detector,this paper proposes two modified algorithms according to the characteristic of the video surveillance.1)Pedestrian Detection Based on Modified Dynamic Background Using Gaussian Mixture Models and HOG-SVM Detection,first,the background modeling using mixture of Gaussians is used to extract the moving objects in the video.Then,through the algorithm of removing shadow,eroding and dilating and Border scaling,our technique make further changes to the foreground of extraction.At the same time,our experiments,based on the INRIA dataset,calculate the Histogram of Oriented Gradient feature of the whole pedestrian and classify by support vector machines classifier.Our experiments indicate that the foreground extracted by the background modeling using mixture of Gaussians can contain the whole moving objects well through the method of removing shadow and border scaling.So the proposed method outperforms the traditional HOG+SVM method in both recognition accuracy and processing speed.2)Fast Pedestrian Detection and Tracking Based on Vibe Combined HOG+SVM Scheme,the preprocessing steps of this algorithm are the same as the above algorithm,the difference is that the last step of this algorithm is using a template matching technique to further track detected pedestrians.Experimental results indicate that the proposed method outperforms the traditional HOG+SVM and GMM+HOG+SVM algorithms in terms of both recognition accuracy and processing speed.