Registration Algorithm Optimization Study Of 3D Reconstruction Based On Kinect

Three-dimensional (3D) reconstruction technology is one of the popular research which focus on the computer vision, reverse engineering and other fields in recent years, and it is widely used in digital preservation of cultural relics, action production and other areas. With the development of computer technology, more and more related industry needs.3D reconstruction which based on a large professional laser rangefinder, the equipment is expensive and complicated to operate. Therefore, the scope of application is not only limited to a handful of basic research in the field, but also limited to the expansion and development of related topics in some ways. With the innovation of technology, Microsoft launched the Kinect, which is a 3D camera, it can quickly and easily complete 3D reconstruction. Therefore, according to the characteristics of Kinect, registration algorithm optimization of 3D reconstruction based on the Kinect has better theoretical significance and practical value.The main contents of this dissertation are as follows:1. We proceed from the structure of Kinect and depth image acquisition principle and analyze the generated reasons for a lot of noise, especially chunk "black hole" region in obtaining the depth image process, in contrast to the common image segmentation algorithm. The gray value of depth map is continuously and does not change significantly, the edge of a large amount of noise, etc. An improved algorithm is introduced into repair depth image. Color image sharp edges characteristic is fully utilized, the edge information of the object to be retained better.2. We analyze the point cloud registration principle, aiming to traditional ICP algorithm which requires good initial values, slow, easy to fall into local optimal solution, the registration process is divided into global registration and partial registration. We use the 4PSC algorithm as a global registration algorithm to provide a good initial value for the partial registration. Then the traditional ICP algorithm is analyzed in details, we propose an improved ICP algorithm to enhance the efficiency of registration.3. We use the MATLAB to do some simulation experiments in the hardware support of Intel (R) Core (R) CPU,8GB of memory. The experimental results show that repair algorithm and registration algorithm proposed in this dissertation to improve the quality of the depth image, enhance point cloud registration accuracy and speed. Applying it in 3D reconstruction can achieve good results.The main innovation of this dissertation in two aspects:1. We find the implicit edge information by an over-segmentation of the color image instead of directly with common methods. The general purpose of image segmentation is divided into meaningful regions based on edge information, and over this region segmentation will be further divided. Particularly, the divided areas had been needed to represent a color image edge. Then divided region is projected to a pretreatment depth image of inter-filtered to achieve the depth image over-segmentation, and hence to retain the edge information of objects in the depth image better while realizing the filter.2. We propose a new threshold at the point of removal stage of ICP algorithm, so when the iterative calculation is needed to computer, we can significantly reduce the number of error points, the algorithm computation time can be optimized.