3D Point-Cloud Surface Reconstruction by Using KD-Tree-Based Mean Shift Algorithm

• Main Authors: Yen-WenWang, 王彥文
• Other Authors: Jenn-Jier Lien
• Format: Others
• Language: en_US
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/57728763484312330915

碩士 === 國立成功大學 === 資訊工程學系 === 104 === This thesis presents a system to mainly process 3D surface reconstruction of scroll. Mainly process noise, and other thesis processing method is removing noise, but our method is noise reservation and reduction that can reserve truth of the information. The system consists of two subsystems: 1) 3D structured light reconstruction system 2) 3D Point-Cloud Surface Reconstruction by Using KD-Tree-Based Mean Shift Algorithm. The first subsystem will reconstruct 3D object surface of point cloud using structured light. Before reconstruction, DLP and camera parameters need to be calibrated by using OpenCV algorithms. After Calibration, we can obtain relationship between DLP and camera. DLP projects gray code pattern onto the surface of the object and decode gray code pattern to obtain the corresponding point between DLP and camera, then it can reconstruct 3D surface of the object. The second subsystem will be case of scroll reconstruction. According to the result observation to segment surface reconstruction and process separation of the upper and lower scroll respectively. Firstly, generate gaussian model and select adaptive standard deviations based on processing result observation to extract outlier noise. It will reserve extracting outlier noise. Next, it will build a KD-tree for point cloud to be able to quickly find K nearest neighbor points, and detect distribution density around the point based on radius. Low density point will be considered as low density noise and be extracted. It will reserve extracting low density noise. Then use mean shift algorithm to push extracting noise back on surface of the object. Finally, according to corresponding point between 2D DLP image and 3D point cloud generate depth Z map based on normalization of depth, and then scan depth Z map to fill the hole. Depth Z map will be returned to 3D point cloud.