Research On Suspicious Person Detection Method Under Low Resolution Video Surveillance System

Intelligent video surveillance system has caused widespread attention anddeveloped a lot. However, video-based human behavior analysis is still complex studyarea. It is an important branch to detect suspicious persons in community and officebuilding for video surveillance. This paper puts forward a new method to recognizethe suspicious personnel based on existing computer vision technology.In general，for a people in the laboratory building, the normal behavior pattern isthat walking from the gate to his office and then goes in and out. However, thesuspicious persons will appear in the office door of every floor. According toprobability statistics, we can define the suspicious personnel’s behavior mode is onesame person who was founded in different locations in the different floors in a certainperiod of time. In this paper, we use the camera of entrance to collect pedestrianimages that enter the building and establish the samplings dataset. Then we collect thedetected pedestrian images for cameras in other floors and compare with the samplingdataset, and then we can detect the locations and times of the same person appearingin. We can determine the person as a suspect if he occurs frequently.The videos we used are taken from the surveillance equipment in laboratorybuilding. We first use the Adaptive Gaussian Mixture Model for backgroundsubtraction to extract the moving objects. Then morphological operations are appliedfor removing noise, cavity filling to extract the main moving object. In the actualmonitoring scene, the moving objects may contain highlights, wall and shadows, sothis paper adopts the method of Histograms of Oriented Gradient (HOG) descriptorsto extract the pedestrians and obtain the position of the pedestrians in the image. Dueto the lower resolution and quality of the video surveillance equipment, the outline ofextracted pedestrian is not accurate. So we use GrabCut to extract the more correct pedestrian from the background. We use the color histogram as the feature of everypedestrian and apply SVM to train the sampling data at the entrance and establish thedataset. After that, the pedestrians at the other floors match with the pedestrians in thedataset.Our tests show the validity of the method which could extract and classify thepedestrian. This method can distinct different people and detect the locations andtimes who appears, so the method can be used to detect suspicious personnel.