Research On Key Technologies In Binocular Stereo Vision Processing

Stereo vision is an important and hot research issue in computer vision, it consists of left and right cameras which capture two images of one scene, by calculating the disparity from the responding pixels of a spatial point in left and right images, then on the basis of the disparity principle and camera parameters, we can get the 3D coordinate value of the spatial point. Stereo vision stimulates human vision system to handle views and get the perception of the outside world, it has important theoretical and practical values, researchers all around the world have done a lot of researches into the related problems, but since this issue is essentially difficult, the application research of stereo vision in industry is still at the initial stage.This research issue aims at applying stereo vision in practical use, especially for the fast visual navigation under normal conditions. The critical techniques of stereo vision are depth perception and 3D reconstruction, this article will carry out a deep re-search into the important and difficult problems in stereo vision, such as stereo match-ing, object segmentation and object reconstruction. The main works comprise the next three contents.(1) Stereo matching is the core research issue in stereo vision, it is the key tech-nique to transfer 2D media into 3D media. The current fast local stereo matching methods can not handle the discontinuous disparities near objects boundaries and the disparity areas formed by initial outliers. To solve this problem, this paper proposes an effective local stereo matching method by extended triangular interpolation. In this method, the reliability of the disparity triangular plane and how to quantify the relia-bility are discussed. According to the reliability of the triangle and the local disparity smoothness, we formulate a Bayesian stereo matching model. The experiment results show that the proposed method can well deal with the objects’ boundaries, and the time efficiency can also meet the requirements of fast visual navigation.(2) Stereo object segmentation is the basis of stereo image analysis and scene recognition. Object segmentation is a basic and difficult problem in image processing, traditional 2D image segmentation methods mainly depend on the basic information (gray, RGB, edge texture) to do segmentation, and the segmentation results have lim-itation. To handle this problem, this paper proposes a multiple-object segmentation method which is based on the disparity from stereo matching. The experiment results prove that, the proposed method performs better than the traditional 2D image seg-mentation methods, and the time efficiency is high, about milliseconds to process one image and it can satisfy the demands of fast object segmentation in visual navigation.(3) The purpose of object 3D reconstruction is to acquire the 3D shape and the space location of objects, the key technique is to get the objects 3D point cloud. To match the same object segments from different views, this paper utilizes two methods: local key points matching and global object areas matching. After the matching stage, we combine the matched objects from different views into the object 3D cloud, then we use OpenGL to display the 3D object. Finally, we implement a prototype system which has realized the related techniques of stereo vision for visual navigation.The core technologies of stereo vision are depth perception and 3D reconstruc-tion, this article carries out research into the related techniques of scene analysis in visual navigation, the specific contents include stereo matching, object segmentation and object reconstruction, we will give our solution for every subproblem, the relat-ed experiments testify that, the research contents in this article have some practical availability.