A Method Of Point Cloud Model Reconstruction Based On Multi Calibrated Depth Camera

The3D scene reconstruction is a part of3D reconstruction, which is a technical ofmathematical modeling for scene. The existing3D scene reconstruction bases on RGB model,through the analysis of the background image segmentation and object modeling, and the generalalgorithm is inefficient. The extension of depth camera provides more and more methods torealize the3D scene reconstruction and provides the basis of RGBD model reconstruction.The paper proposes a reconstruction technique of point cloud model based on multicalibrated depth camera. At first, the method should calibrate depth cameras and use the camerasto obtain a series of dense unorganized point clouds which describe the information of the objectsbeing measured in the scene by scanning the scene from different perspective; Then, the methodgets the initial point cloud by preprocessing the dense unorganized point clouds and using themotion parameters of the depth cameras on the point clouds. In the end, the method takes thepreprocessed point clouds as the initial registration to do the precise registration for completingRGBD model reconstruction. The main research is following:(1) Chapter2designs a system framework for point cloud model reconstruction. The systemincludes data acquisition platform and algorithm implementation platform. In the data acquisition,the paper uses two Kinects to collect scene data with Kinect Fusion simultaneously. Theexperiment scene is a corner of lab which has a breadth of2~4m, and the Kinects are put1.5maway from the scene, which are set the vertical angle to0and have a150~170angle betweeneach other.. The algorithm implementation platform programmed with PCL for camera calibrationand point cloud processing.(2) Chapter3shows the research on depth camera calibration. The paper uses the SUSANalgorithm to detect the angular point in the images, which has high anti-noise ability, range ofapplication and high accuracy. Then, it uses chessboard calibration algorithm to calibrate twoKinects and the relative position of each other.(3) Chapter4shows the research on point cloud pretreatment of depth camera. This paperuses the bilateral filtering to remove the large-scale noise from the dense unorganized pointclouds with a lot of noise points and uses triangular mesh models to repair the point cloud modelwith defects. At the same time, the paper compresses the processed point clouds using the octreesupplied by PCL to reduce the storage capacity of point clouds. It also improves the processingspeed of subsequent point clouds.(4) Chapter5proposes a reconstruction technique of point cloud model based on multi calibrated depth camera. This paper studies the classic ICP algorithm and describes a registrationmethod for point clouds obtained by calibrated Kinects based on the shortcoming in the existingICP algorithm. The method calibrates two depth cameras and the relative position of each other.Then, it dose initial registration with the calibration parameters for the preprocessed point clouds.In the end, the method uses the ICP algorithm to do the precise registration.The experimental results show that the method described by this paper has high registrationaccuracy and computational efficiency compared with traditional point cloud registrationalgorithm. It also has many practical applications used in practice with its fast convergent rate andregistration accuracy.