3D Indoor Scene Reconstruction With Kinect Depth Camera

3D reconstruction technology can digitalize objective reality scenes to reappear and process scenes in the computer. Reconstruction information can help us understand the whole environment around us and process actual better conveniently and generally. It is the foundation and condition for mobile robot to do autonomous navigation and path planning. This paper presents the problem of3D indoor scene reconstruction with Kinect depth camera, and we will analysis and solve several problems, such as data acquisition, matching algorithms among multi-scenes, data reduction, et al, so that can realize3D indoor scene reconstruction effectively and fast.Matching among multi-scenes is the core of3D reconstruction, so effective matching algorithm is the important premise of reconstruction, and we solve transformation relationship between scenes essentially in order to achieve coordinate unification under different viewpoints. A full3D reconstruction system that utilizes a joint optimization algorithm combining visual features and depth information is presented. In order to realize multi-scenes matching, we use the singular vector decomposition (SVD) algorithm and the iterative closest point (ICP) algorithm based on SURF features being extracted from original gray images. Some experimental results show SVD algorithm to have an advantage over ICP algorithm in terms of effectiveness and real-time.Because of perspective and distance restrictions, it can emerge accumulative error during3D reconstruction, especially in the position of loop closure. In order to get good results of3D scene reconstruction, pose drift caused by loop closure need to be solved. We divide into loop closure for detecting and processing overlap region. Firstly, we build pose graph between nodes for loop closure detection based on visual overlap so that can process accumulated error, followed by pose graph to achieve globally optimization in order to balance or distribute error. Matching between adjacent scenes is bound to create data redundancy. According to this problem, we propose2D grid structure to process data simplification. Many groups of experimental results of3D indoor reconstruction prove the effectiveness of the algorithm in this paper.