Research On Object Detection Method Of PTZ Camera

Moving object detection is an important topic of computer vision, which playsa very important role in intelligent surveillance and other fields. PTZ camera canmimic the action of gaze, beating, and follow of the eyeball of hunman active vision,expand the monitoring scope and enhance the scene perception. It also canovercome the shortcomings of the narrow view of ordinary camera and the lowresolution of wide-angle camera, so it is widely used in monitoring. The video takenby PTZ camera contains its own motion, which brings difficulty and challenge formoving object detection.This paper firstly summarizes the basic theory of object detection for PTZcamera. We compensate the camera motion through motion estimation. After that weregister the current video frame to the panoramic background image. Then we buildpanoramic background model, detect moving object and update the sub backgroundmodel corresponding to current frame.In order to overcom the accumulative error of the frame to neighbouring frameimage registration method, we study the image registration method of frame toreference frames and put forward the reference frame updating mechanism based onfeature matching. By using the ratio of the feature matching number between thevideo frame and the reference frame to the feature number of this reference frame,we obtain the best match image of the reference image. For the non anchor frame inthe reference frames, we use its best match image’s feature and the feature matchinginformation between its best match image and other reference frames to replace thisnon anchor frame. For the anchor frame in the reference frames, we use its bestmatch image’s feature, the feature matching information between its best matchimage and other reference frames, and the relative rotation matrix between its bestmatch image and the panorama background plane to replace this anchor frame.In order to reduce the effect of the image registration errors, we improve thebackground modeling method based on the traditional Gaussian mixture model(GMM). We propose a novel block-based panoramic background modeling methodbased on Gaussian mixture model. Firstly we find the corresponding point on thepanorama background image to the pixel on the current frame. Then we use the thepixel on the current frame to match the model parameters of the correspondingpoint’s surrounding pixels. Last we use distance-weighted method to determine thatwhether this pixel on the current frame blongs to a moving target or not. Meanwhile,we propose a strategy to update the sub panorama background model based onbackward projection. We do experiments on the actual taken video by PTZ camera. The method inthis paper can effectively detect moving objects. The result verifies that our objectdetection method is reasonable and the improved method is effective.