Research On Key Technologies Of3D Reconstruction Based On Multi-plane Detection In Point Clouds

3D reconstruction plays a very important role in economy, life and military field, such as intelligent city, navigation and positioning, battle command, which has always been the hot spot of the scientific researchers to discuss and research over the years.3D reconstruction consists a series of links like camera calibration, point cloud data acquisition, model fitting, scene rendering and so on. Because planes take a big amount in urban scene and have rich semantic information, automatic fitting plane models in real-time has a great significance for3D reconstruction, so this topic focus on improving the plane detection algorithm in point clouds, and reveal the results by3D visualization rendering, before that, the Kinect is used to collect the registration of3D point clouds and texture image data.In this paper, we first expound the principle of the Kinect and pinhole camera imaging model, in-depth study and set up the Kinect color camera and depth camera distortion model, calibrate Kinect by checkerboard through color and depth, moreover, according to the principle of perspective imaging, the method to obtain registration of point clouds and color image data is proposed.Based on this, this paper proposed a step by step strategy to detect multi planes. The method Put forward the concept of the point cloud density change rate, according to build volume density gradient histogram roughly segment point cloud, and then use the Multi-RANSAC to detect the initial planes which decreased the data amount to be processed in the loop iteration process and eliminate the accumulated noise on the accuracy of algorithm. To the detected initial planes, according to the relation of detected initial planes’normal vector and the center of mass position, this paper proposes a new merger constraint to optimize a plane, which solved the problem of the different depth parallel planar chaos to merge, improves the integrity of the fitting model.Finally, based on texture image achieved by Kinect in the front and the result of plane detection, according to the classic volume package model algorithm, the method of convex hull seeking was improved by deriving from the world coordinate space to the2D image space, which reduced the amount of calculation of the algorithm. The simple scenario model rendering, reconstruction of texture mapped scene as well as specific interactive control function were completed by using OPENGL programming, by which the scene rendering speed was improvedOur algorithm is implemented under Visual Studio2008programming environment, using C++programming language combined with OPenCV2.1, OpenGL.