Matching Technologies For Binocular Stereo Vision Images

In recent years, stereo matching technique is a very important branchin computer vision. Its main purpose is to find relationship betweencorresponding pixels among multiple images according to multiplephotographed images of a scene. Thus the object contour can bedetermined. All3D coordinates of the object can be obtained so the whole3D information of the object can be recovered. This technique is nowwidely used in virtual reality, robot navigation,3D measurement and so on.Nowadays, it is one of the hottest research topics in computer vision.Among all the stereo vision techniques, binocular stereo vision is avery important research point. The basic principle of binocular stereovision is to simulate the procedure of human eye observing3D objectsthrough the establishment of two viewpoints on the observations of thesame object in order to gain perceptual images from different perspectives.Once we have calculated the pixel-level deviation of a certain point on theobject among two images, we can obtain the depth information of thatpoint by the principle of triangulation.A complete binocular vision system includes image acquisition,camera calibration, stereo matching and3D reconstruction. Imageacquisition aims at getting stereo image pair which is taken from the samescene by moving or rotating the camera. The goal of camera calibration isto obtain internal and external parameters. Stereo matching is the key pointof this paper which aims at establishing corresponding relationshipbetween pixels of two2D images, calculating their difference on x-axisand generating disparity map. The final step3D reconstruction calculatesthe depth information of each point using triangulation combined with camera parameters and disparity map thus constructing the3D model ofthe object.In this paper, we survey the existing popular local matchingalgorithms and analyze the advantages and disadvantages of each method.Finally, we combine two methods among these and propose a stereomatching algorithm based on region-adaptive. Our method canautomatically choose the corresponding algorithm according to theproperty of region which the pixel belongs to. The proposed method notonly improves the depth information on object borders but also has a goodaccuracy on low textured region, high textured region and repetitivetextured region. Finally, our method achieves near real-time performancein time efficiency.Our method can be generally divided into the following three steps.First, we do edge extraction and image segmentation on the originalimages. The generated edge images and segmentation images can be usedas a guidance in the following steps. Then, we adopt correspondingmatching method which depends on whether the pixel is an edge pixel.Finally, we present a simple and efficient disparity refinement methodwhich uses the information of near-by pixels to eliminate the noisy pixelsin the disparity map. The final experiment result shows our method has agood speed-accuracy trade-off and wide applications.