Research On The Theory And Algorithms Of Three-dimensional Reconstruction From Infrared Images

Reconstructing3D models from2D images has been a hot research topic in the field ofcomputer vision. It is a comprehensive and cross research area and has a very wide applicationprospects in the fields of visual navigation, historic preservation and medical diagnosis, etc. Forscenes and objects, this paper provides a deep research on the theory and algorithms ofthree-dimensional reconstruction from infrared images.Firstly, discussing a method for reconstruction of the scene, we focus on vehicle infraredimage. Based on the characteristics of the vehicle infrared image, we proposed a3Dreconstruction method of vehicle monocular infrared image, which combines super pixelssegmentation and PP-MRF model. The PP-MRF model has been already applied to the3Dreconstruction in the visible light area successfully. We improved the PP-MRF model according tothe characteristics of the vehicle infrared image and make it emphasize the analysis of big panelstructure such as road and sky in the image. Such improvement can implement3D reconstructionof the vehicle infrared image with better result. Firstly, we segment the image into a series ofsmall areas where the texture and brightness are similar. These small areas are called super-pixels.Then the PP-MRF model is trained, which can analyze the panel parameters and estimate thedepth of each super-pixel of the testing image. The experimental results show that the proposedmethod can estimate the depth value of vehicle monocular infrared well and it can rebuild the3Dscene properly.Secondly, discussing a method for reconstruction of the object, in order to obtain thecomplete three-dimensional reconstruction of object, we need to take the object from multipleangles. The reconstruction process can divided into two parts, the first part is obtaining the depthinformation based on infrared images, and the second part is used the depth data to reconstruction.In this paper, we mainly discuss about the second part. Firstly, we introduced the workingprinciple of Kinect, and use it to detect the depth of the object. Secondly, according to the imaging principle of the camera and the calibration technique, we obtained the three-dimensional pointcloud data by the depth data. Thirdly, on the base of the classical ICP algorithm, we proposed aregister algorithm for the3D point clouds data from global registration and local registration. Inorder to achieve a global registration, we selected Spin Image method, and then we used animproved ICP algorithm for local registration. Finally, for meshing the point cloud, the paperdescribed the Power Crust algorithm briefly. The experimental results demonstrates that3Dmodels can be reconstructed, which verifying the feasibility and efficiency of the algorithmsmentioned in this paper.