Research And Application Of Depth Image Restoration Technology Based On Stereo Vision

As an important branch of computer vision,binocular stereoscopic vision obtains the depth information of the object to be observed by calculating the parallax on the left and right view of the binocular camera by calculating the point in the 3D world.As the binocular stereoscopic vision is compared with other ranging methods Non-contact and strong robustness,so binocular stereoscopic vision in the automatic driving,behavior detection and identification,virtual reality,security and other fields have a wide range of applications.With the continuous development of binocular stereoscopic vision,but some problems are still not resolved,and in some applications,binocular vision has a strong applicability,so this paper is the main study of binocular stereo vision matching method and related Application,mainly on the binocular stereo vision in the image matching,coordinate transformation,depth map applications and other aspects of the theoretical practice and technology research.In order to make the local stereo matching algorithm have higher matching precision and faster matching speed,we propose a seed point propagation algorithm based on the combination of feature seed point and regional seed point.The FAST feature is extracted and then the descriptor is used as the characteristic of the ORB.The excellent feature is chosen as the characteristic seed point,and the similarity between the matching points is matched by the AD-Census.The Canny edge detection is used to extract the dynamic detection window to obtain a more accurate match result in the low texture region.Then,the regional seed spots are screened according to the principle of left and right consistency,and the dense parallax is obtained by propagating the seed points and the characteristic seed points.Finally,the parallax is optimized by median filtering and subpixel refinement.A matching algorithm with low matching speed and low error is obtained.According to the RANSAC algorithm,the plane information in the depth imageis extracted,and then the plane points in the depth map are removed according to the fitted planar equation.The use of face detection or head and shoulder detection to get the location of the human body information,so as to extract the contours of the human body.The motion history energy map of three directions is put into the convolution neural network as the three channels of the color image,and finally the appropriate classifier is obtained.The accuracy of motion recognition is increased from 72% to 90%.