Scene 3D Reconstruction By Binocular Stereo Vision And Kinect

Stereo vision is a common method to get the depth of scene, which can obtain a large range of depth through different configuration of cameras' baseline and focal length, but is only suitable for texture rich and bright enough scene. Kinect is a kind of depth camera that can get the depth of scene by using infrared active structure light, so Kinect can reconstruct the texture less regions and the darker scenario in a scene. However, Kinect has a limited scope of depth. In order to combine the respective advantages of these two complementary imaging technology, this paper carried out the following work:Firstly, use the binocular stereo camera and Kinect for indoor scene 3D reconstruction separately. In this paper, 3D scene reconstruction based on binocular stereo vision fallow such steps: off-line calibration, on-line correction, dense matching, obtain scene's disparity image, generating 3D points cloud and something else for points cloud process. For Kinect indoor scene reconstruction, this paper firstly calibrate the depth camera, and then convert its depth image to the 3D points cloud.Secondly, combine binocular cameras and Kinect together to get a more complete and accurate 3D points cloud model of the indoor scene. Because the roughly iterative nearest neighbor algorithm(ICP) cloud not align the two points cloud recovered by binocular cameras and Kinect, this paper design a "two-step" scheme, which the FAST feature points based Random Sample And Consensus(RANSAC) initial registration carried out firstly and LM_ICP algorithm based accurate registration latterly. The experiment show this scheme can achieve a good balance between time efficiency and registration performance.To verify the performance of the 3D reconstruction of indoor scene by the combination of binocular cameras and Kinect imaging system, this paper use the combination reconstructed two kinds of scene: stereo vision and Kinect are not suitable alone. They are texture less region, dark environment, repeat pattern place, infrared light absorbing and reflecting environment, depth is too far or too near place. The test results show that the precision and integrity of the 3D points cloud both have some improvement against the single imaging mode's result.