Indoor Scene Reconstruction Technology Based On Kinect

In recent years, Computer Vision has developed rapidly. As an essential part of it, 3D reconstruction draws more and more people’s attention. Through multiple images, it reconstructs a 3D object model or a 3D scene. Currently, 3D reconstruction technology has been widely used in various simulation scenes and intelligent systems, such as virtual reality,robot navigation, and modern medicine. There is an increasing demand for simple, flexible and real-time acquisition procedures.In this thesis,with the help of Microsoft’s Kinect device,we obtain depth information of indoor scene, and then achieve the indoor scene reconstruction by SLAM (Simultaneous Localization and Mapping) algorithm. The followings are the main research work of this thesis:Firstly, Kinect calibration. First of all, the article analyzed the principle of achieving depth information and the principle of the camera imaging. At the same time, it also introduce the model of camera distortion and geometry relationship of two cameras. After the Kinect calibration, wonderful effect is achieved.Secondly, Researches on acquiring point cloud data. The paper analyze and compare the two methods (one is based on OpenCV and OpenGL, and the other is based on PCL). The first method read RGB image and depth image of Kinect by OpenCV and generate the point cloud data by defining the point cloud data structure. And then utilize OpenGL to show the indoor scene and realize the operation of zoom-in, zoom-out, and mouseDrag. Point cloud data is obtained with another method by PCL. By comparing the two methods,we selected the first one.Thirdly, Feature detection and match. This paper researches the main feature detection and descriptor algorithms, such as SIFT,SURF,FAST,FREAK,ORB and KAZE. And by taking advantage of RGB-D information the accuracy of feature detection and match is improved in 3D reconstruction. With the comparison of the several algorithms in real-time and robustness I find the method using FAST detector and FREAK descriptor is superior to the others. So I select it into RGB-D SLAM. And at last I study the Stitching technology.Firstly estimate the Kinect Camera pose, then we stitch point clouds with transforming into PointCloud function of PCL.Finally, The loop closing detection based on G20. Due to the environment feature problems, it beeps on error easily in the match of the current frame and the last, which always lead to the error accumulation. By researching the G20 algorithm and loop closing detection,we find it can reduce the error accumulation efficiently. And this paper bring the G20 and loop closing detection in RGB-D SLAM system.