Research On Detection System Of Moving Object Based On Binocular Vision

Detection system for moving object based on binocular vision is a hot issue in current researching of computer vision field. Compared with monocular vision, real information of three-dimensional scene can be restored and the extent of the differences in the distance about targets can be reflected through binocular vision. So it can be widely used in the photogrammetric, robot, intelligent monitor and many other various vision systems, and it has the important theoretical researching value and application significance.The research status and the development trends of the camera calibration, feature extraction and matching, moving object detection related to binocular vision are described in this paper, and the implementations of target detection about monocular vision and binocular vision are introduced in detail. The method of target recognition is analyzed comparatively, and the classifications, the extraction of image features, the image segmentation are included. And then the basic principles of stereo vision and the measurement processes are described. Calibration method of camera is introduced, and the theory, the model of imaging about camera, the calibration method, and the parameters adjustment are included. A high-precision method used for camera calibration in this design process is proposed. The grid corner of each image is extracted firstly, then the calibration parameter of intrinsic and extrinsic is computed by minimizing the re-projection error through gradient descent, and the distortion coefficients of lens are adjusted. In the monocular visual experiment, an algorithm based on the in memory of real-time adaptive background template library updating and creating is proposed. The template library is initialized and updated Firstly, and then the targets are detected by the method of sparse matrix and full matrix. And the pending targets are determined and extracted by the adaptive method. In the technologies based on binocular vision, disparity maps that acquired through sequential images from the left and right cameras are matched exactly, foreground images are obtained through the two-way differential multiplication method for the disparity maps. The foreground objects are detected according to the disparity maps rather than the color images directly obtained cameras. The disparity map will be divided into several sub-images before the detection, and it will be marked for the largest connected region. The moving targets are identified and tracked through the location with hierarchical strategy according to the information of depth which is the depth of field contained in the foreground image.The proposed algorithms are studied by simulation using Matlab software tools. Experimental results show that:(1) By the moving targets detection method based on monocular vision proposed in this paper, illumination changes, background interferences, occlusions and many other noises can be effectively overcame, and the moving targets can be real-time tracked accurately.(2) By the binocular vision-based moving object detection method proposed in this paper, the several different types of moving objects can be detected in real time and accurately, and the desired results are achieved, so it has the practical meaning of application.