Moving Objects Detection Based On Improved Gaussian Mixture Model

Moving object detection is the foundation of tracking, recognition and behavior understanding in the field of computer vision and intelligent surveillance. This thesis has been made a research on moving object detection based on Gaussian Mixture Model (GMM) using fixed-camera. The main works of this thesis are as follows:(1)This thesis analyzes and compares moving object detection algorithms, including sequences frames subtraction, optical flow method and background subtraction. This thesis also elaborates on the advantages and weaknesses of typical background models from computation, detection effect, anti-interference ability and memory requirement.(2) The resolution of each frame is resized by bilinear interpolation. Impulse noise and salt and pepper noise, which badly influence GMM, are restrained by 3-D vector median filter. Experiments indicate that this algorithm gain better PSNR than scalar median filter and 2-D vector median filter.(3)A moving object detection algorithm based on improved GMM for fixed-camera is proposed. The algorithm detects foreground by GMM and then classifies it into real changed area and false alarm area. The model sets the update rate in false alarm area larger than other area to solve "ghost" and sudden illumination change problems. Experimental results indicate that this algorithm responds real scenes quickly, and segment moving objects accurately.(4)After moving object detection, isolated foreground points are restrained by median filter and hollow-out object is filled by morphology filter. Shadow detection based on HSV and rgb color conversion is carried out. Experiments indicate that these algorithms can effectively remove the shadow.(5)The result of moving object detection is binary image sequence. After setting a virtual detection area perpendicular to the direction of vehicle flow in appropriate place of driveway, the system count foreground points to judge whether there is a car in the virtual area. So vehicle flow can be counted by monitor the state of the virtual area.