Indoor 3D Reconstruction Based On RGB-D Camera

In computer vision and computer graphics,research on 3D reconstruction has always been a hot topic,and it is moving forward with the goal of low cost,fastness and precision.In 2010,the Kinect series of depth cameras developed by Microsoft greatly promoted the research of low-cost 3D reconstruction.This paper is about the 3D reconstruction of Kinect v2,the second generation depth camera of Kinect series.The main research contents are as follows:(1)The working principle and components of Kinect v2 camera are studied and introduced in detail.The stability of Kinect v2 measurement output is experimented.A fan is added to the outlet behind the original Kinect v2 hardware.The experiment shows that the output of Kinect v2 is more stable and the standard deviation of output is greatly reduced after the external fan is added.The Kinect v2 camera is calibrated by Zhang Zhengyou's checkerboard calibration method.The internal parameter matrix of color camera and infrared sensor and the external parameter matrix of color camera relative to infrared sensor are obtained.(2)Aiming at the noise problem of the depth image of Kinect v2,an improved bilateral filtering is proposed.A binary function is used to replace the gray-scale weight coefficient in the bilateral filtering algorithm,which greatly improves the computational efficiency and combines the color image with depth.The holes in the image are filled.After that,the two are combined to generate the three-dimensional point cloud of the scene,the point cloud is stored in the structure of kd-tree,and a series of preprocessing operations are carried out on the point cloud,including point cloud simplification,removal of outliers and estimation of normal vectors.(3)For the 3D reconstruction under continuous video stream,the point cloud needs to be registered under different perspectives.Here,a two-stage point cloud registration algorithm based on ISS key points is proposed,which divides the whole registration process into initial coarse Registration phase and precise registration phase.In the first stage,the intrinsic shape signature of the point cloud is used to extract the key points,and then only the FPFH(Fast Point Feature Histograms)neighborhood features at the key points are calculated,then,the point cloud initial transformation matrix is calculated by the sampling consistency registration algorithm,and the point cloud is initially registered.In the second phase,the improved iterative closest point algorithm is used to complete the accurate registration work.Finally,the complete point cloud model after registration work is reconstructed by the greedy projection triangulation algorithm.