| 3D reconstruction of target and scene is an important part of computer vision，andit is also the research focus.3D reconstruction based on binocular vision is thetechnology of obtaining the3D information of target object from2D image sequence.Although the reconstruction accuracy is not high, and the technology is difficult, itpossesses the advantages of low cost, easy application, low environmental requirements,and it can be widely used for virtual reality, robot navigation, industrial inspection,geological survey and object recognition. This paper has a depth research on obstaclerecognition and reconstruction based on binocular vision, and mainly completed thefollowing work:1. On the basis of reading a large number of domestic and foreign relatedliteratures about obstacle recognition and reconstruction based on binocular vision,summarized the research status.2. Introduced the related basic knowledge about obstacle recognition andreconstruction based on binocular vision, including the principle of binocular vision, thedisparity map and depth map, three coordinate systems, camera model, epipolargeometry, fundamental matrix, essential matrix, quasi-dense matching, etc.3. Based on the studying of the popular method of obstacle avoidance, proposed anobstacle recognition method which based on the combination of color segmentation andheight detection. In this method, firstly, segment the binocular image sequence, extractthe obstacle area. Then do some processing, such as feature extraction, feature matching,quasi-dense matching and so on, calculate the obstacle’s3D information in the image.Finally, calculate the true height of obstacle according to the height detection method,and restore the true3D information of the obstacle.4. In order to improve the accuracy of obstacle reconstruction, design andimplement a quasi-dense matching method based on the best seed sorting. Firstly, sortthe seed region based on the matched feature points. Secondly, from the best seedregion, spread the matching relations to the whole image through a certain regiongrowth method. Finally, using RANSAC method to eliminate wrong matching, andobtain the final quasi-dense matching.5. According to previous theory and derivation, analyze the workflow and module functions of the whole system, and design an obstacle reconstruction software system,then, through experiments to verify the feasibility and correctness of this method. |
