Research On Indoor Scene Reconstruction

In the area of image processing and computer vision,3D reconstruction forvariety scenes is always a significant and valuable subject. Compared with3Dreconstruction of other scenes,3D reconstruction for indoor scenes could be appliedinto some dangerous yet indispensable activities (rescue the fire, exploration of backrooms, and archaeological excavation, etc). However, due to the occlusions andillumination variations in indoor scenes, recovery the3D structure of indoor scenesremains a challenge problem. To the traditional methods, it is difficult to recover the3D structure from the distorted image with many occlusions, derived from theocclusions and illumination variations. Nevertheless, we propose a novelty algorithm,which recovers the frame structure from a monocular image, then renders therecovered frame with the revised monocular image. Beside traditional methods, whichutilize image sequence, our method, which based on monocular image, is moreeconomic and efficient. The main contributions and novelties can be summarized asfollows.1. We propose an algorithm based on circular template under image coordinate.We first analyze the gray distribution of the local area of chessboard corner to achievethe properties of convoluted image and define the chessboard corner with thoseproperties. Finally we extract the chessboard corners’ coordinates through thoseproperties and our detected chessboard corners can reach sub-pixel level in merelyone step by employing the image coordinate. Experiment results show that ouralgorithm can achieve better results in both simple backgrounds and complex scenes.Applying our detector to camera calibration, we obtain smaller re-projection error,which is less than0.3pixels.2. A chessboard corner detector based on image physical coordinates and a roundtemplate is proposed. The physical coordinates allowed our detected chessboardcorners to reach the sub-pixel level after only one step. We first covered the distortedchessboard corners by utilizing the morphological dilation. Then, we employed ourround template to pass through the dilated image and ultimately determine thechessboard corner coordinate by analyzing the grey distribution of the traversed round template and calculating the centroid of redundant points. The experimental resultsshow that our algorithm performs better than other algorithms in both simplebackgrounds and complex scenes. By applying our detector to camera calibration, weobtained a smaller re-projection error, thereby proving the validity of our proposeddetector.3. An approach for frame structure recovery based on line segment refinementand voting is proposed. We refined line segments by the revising, connecting, andadding operations. We then propose an iterative voting mechanism for selectingrefined line segments, where a cross ratio constraint is enforced to build crab-likemodels. Our algorithm outperforms state-of-the-art approaches, especially whenconsidering complex indoor scenes.4. We calculate the camera parameters from the coordinates of the chessboardcorners and rectify the distorted image with the camera parameters, then render therecovered frame with the texture from the rectified image. The experimental resultsdemonstrate the validation of the proposed algorithm.