Research On Obstacle Detection Based On Stereo Vision Technique

Stereo Vision is an important research field of Machine Vision, which can obtain much more information comparative to monocular vision, and then, more broad scope of application it can be employed to. In recent years, Stereo Vision has been applied to Obstacle Detection, Virtual Reality, Robot Navigation and so on, among these fields, Obstacle Detection can be applied to intelligent vehicle navigation system, having great significance for the further development of intelligent vehicle driving and mobile robots.In this paper, aiming at improving the effectiveness of Obstacle Detection, in-depth research of key technology of Stereo Vision is conducted, including sparse matching based on feature points and dense matching based on belief propagation.For how to detect obstacles by stereo vision technology, research contents and innovations of this paper include the following aspects:1. Stereo matching based on local robust feature descriptor. To solve the problem of SURF (Speed Up Robust Feature) that it works too slow, a stereo matching method based on local robust feature description is proposed in this paper, which is robust on feature extraction as SURF and efficient on local feature description as DAISY. And considering the characteristic of high dimension of DAISY, random KD tree is adopted for matching. And experiments show that this method has been greatly improved in speed compared with SURF.2. Improved dense matching based on belief propagation. Considering the problem that existing matching algorithms work too slow, a Weighted Average of Color Components Difference is constructed for computing the initial disparity map of image, comparing to SAD(Sum Absolute Differences), the constructed function needs less amount of computation and works more fast. And a belief propagation method combing color segmentation is proposed for optimizing the processing of the disparity map. The experiments show that the proposed methods obtain similar results with existing methods, while the matching speed has improved greatly.3. Obstacle Detection. First, V-disparity map can be generated by the resulting dense disparity of Chapter IV. Second, Line extraction can be carried out by Hough method on V-disparity map. Finally, according to the performance forms of obstacle plane in V-disparity map, obstacle detection can be performed by analyzing the lines in V-disparity map.