Research On Algorithm Of 3D Reconstruction Based On Stereo Vision

Three-dimensional reconstruction based on stereo vision is a core subject in computer vision, it also plays a very important role in the areas of robot navigation, virtual reality, architecture and industrial design. At present most research work focused on three-dimensional modeling of a single object or rebuilding sparse feature points, in this paper, we propose a three-dimensional reconstruction methods for large-scale scenes. Firstly, we take consecutive frames of a scene with a binocular stereo camera. Secondly, a visual odometer is used to estimate the camera motion between consecutive frames, a disparity image is computed from the stereo images and is used to produce point cloud data in the camera coordinate frame. Finally, the camera pose with respect to a fixed global coordinate system is then used to register the point cloud data from each frame to the global coordinate system, and then we have obtained consistent three-dimensional point cloud data of the whole scene. This paper mainly has the following several aspects of research work.(1)A new unit quaternion-based visual odometry is presented for computing the motion of the camera. Scale invariant features in consecutive stereo frames are extracted, and then these features are matched and tracked by a circle matching procedure. The unit quaternion-based method is employed to compute the rotation matrix and the translation vector between matched features in two consecutive frames. In order to obtain a more precise transformation parameters, we iteratively minimize the sum of re-projection errors. The experimental results on indoor and outdoor data set show that our proposed algorithm is more robust and achieves a higher accuracy compared with some existing algorithms.(2)Stereo matching is used to recover the depth of scene through computing disparity between matching pixels in the same scene under different viewpoints. A dual segmentation based stereo matching algorithm is presented by improving the traditional approaches. Firstly, it is to under-segment the reference image so that each region in the image contains sufficient cues for plane fitting. Secondly, it is to segment the initial disparity map for detecting and re-segmenting those under-segmented regions in reference image, and the processed results are plane fitted. Finally, to minimize processing time, only invalid regions are iteratively optimized by an inter-regional cooperative procedure. The experimental results on standard test set and real data show that our proposed algorithm achieves higher matching accuracy with less time than other well-established algorithms.(3)We can obtain three-dimensional point cloud of the whole scene by transform each frame’s 3d points to a global world coordinate system according to the estimated camera motion. We present a key-frame-based data fusion method for dealing with the problem of storage requirements grow rapidly. Firstly, detecting key frames in the frame sequence, and then fuse key frame’s point cloud, to remove redundant further, we use re-project method to detect overlap areas between adjacent key frames and keep their three-dimensional mean. Experimental results show that our approach can reconstruct scene’s 3d point cloud efficiently and accuratly, with less redundant data.