The Design And Implementation Of Obstacles Detection Algorithm By 3-d Reconstruction Based On Monocular Vision

Vision-based navigation and 3-D reconstruction, which have great value in research and application, are the key areas of the computer vision research. For obtaining the 3-D coordinate of the background from the 2-D images captured in a scene by two or more cameras, the stereo vision detects the obstacles by the scene interrupted, understood and navigated using the parallax information and the external parameters of the cameras. The results of stereo vision and 3-D reconstruction are more accurate because of the precise equipment, however, there are some limitations like higher cost, lower effective, larger computational complexity and higher precise image registration and so on. Once some mechanical failure happened on one of stereo cameras, the system cannot work naturally. And this paper has proposed a method which can detect the obstacle on the path of the vehicle based on 3-D reconstruction derive feature points matching of sequence images captured from monocular. The principle of the algorithm is to recover the motion parameter of the camera by the parallax-images, then get the depth information of the feature points in the images and cluster of them. So we can implement the purpose of objects detection. The algorithm has some advantages, for examples, little computational complexity, lower cost and flexible.There are some technology about 3-D reconstruction based on sequence images which have been calibrated:feature points detection and matching, the estimate of fundamental matrix, the estimate of the ego-motion of the camera, Euclidean reconstruction and so on. Feature points detection and matching and fundamental matrix play an important part in the 3-D reconstruction. Based on above, the paper has discussed as follows:(1) Introduced the theories of epipolar geometry, fundamental matrix and the function of them in vision navigation and 3-D reconstruction.(2) Studied the feature extracting and feature matching. Emphasized on the algorithm of Lucas-Kanade tracking and improved on it and we have a good result of road images. (3) Studied the classical methods and robust methods for solving fundamental matrix like RANSAC and proposed some amelioration on RANSAC. As a result, the accuracy of fundamental matrix have been improved.(4) Recovered the real motion parameters of the camera includes rotation matrix R andtranslation vector T and resumed the 3-D coordinate of all the feature points in the image.(5) Studied the algorithm of clustering. Proposed a method based on grid clustering and hierarchical clustering cluster all the 3-D points and detected the obstacles in the image.In the end, the paper provided the analysis of the experimental results and problems of obstacles detection based on 3-D reconstruction of monocular, showed some results of obstacle detection.