Research On Three-Dimensional Reconstruction Based On Kinect

With the rapid development of digital technology,3D reconstruction has become one of the research hotspots in cutting-edge areas like computer vision, computer graphics and artificial intelligence. However, traditional 3D reconstruction devices are not perfect in terms of cost and operational flexibility.Kinect is developed as a gaming peripheral for Xbox 360 by Microsoft, but provides a cost-effective solution for 3D reconstruction unexpectedly, which has caught the attention of researchers. Around the topic of 3D reconstruction based on Kinect, the main achievements of this thesis include:(1) The structure of Kinect and the operating principle of achieving depth data were studied. OpenNI SDK was also introduced to make sure the achievement could be used with the latest depth camera. A method of checkerboard calibration for Kinect was reported, which was applied to both the color camera and the depth camera with shielding infrared transmitter. The intrinsic parameter matrix and relative position of two cameras were calculated to prepare for generating point cloud data.(2) The quality of depth image obtained by Kinect is easy to be decreased due to the instability of depth data. An improved algorithm for depth image restoration was presented, which firstly completed hole-filling in spatial-temporal domain, and then smoothed the image using improved joint bilateral filter with a threshold. The corresponding color image was aligned to the restored depth image to generate point cloud, and the acquired 3D point clouds were organized with k-dimensional tree and pre-processed for reduction, outlier removal, normal estimation and segmentation.(3) A two-phase registration algorithm for point cloud based on neighborhood feature was proposed for aligning point clouds obtained under different views, which includes initial alignment and accurate registration. The algorithm firstly extracted the key points of point cloud according to intrinsic shape signature (ISS), and then computed fast point feature histogram (FPFH) based on the neighborhood feature of key points which decreased the computational expense compared to the whole points. Sample consensus initial alignment (SAC-IA) algorithm was used to obtain the initial point cloud coordinate transform matrix. During the phase of accurate registration, improved iterative closest point (ICP) algorithm should be applied to finish the second mosaicking. The registration experimental results showed that the improved algorithm was efficient and the obtained all-in-one point cloud was transformed into mesh model with greedy projection triangulation (GPT).(4) From reading data streams of Kinect to generating 3D models, the whole process was implemented by programming in the Visual Studio 2010 development environment with OpenNI, OpenCV and point cloud library (PCL), which can be provided as a platform for further researches such as virtual reality and visual recognition.