Research On Parallel Algorithm Of Image-based 3D Large-Scale Scene Reconstruction

The technology of image-based large-scale 3D scene reconstruction is the prerequisite for large scene visual ability of computer, and the research has great academic value in computer vision. Furthermore, it has a wide range of practical applications in mapping and urban planning, heritage, military,3D entertainment and many other aspects. Therefore, the technology receives wide attentions from scholars at home and abroad. However, the current reconstruction algorithm is not applicable for 3D large-scale scene reconstruction. The problem of these algorithms is that if the image has large radial and affine transformation, it cannot detect large and stable characteristic points. At the same time, the optimizing accuracy is not high enough when bundle adjust algorithm optimizes the final result of the 3D reconstruction. These algorithms cost tremendous time to process image feature extraction, feature matching and bundle adjustment. In response to these problems, this paper proposes a radial-distortion and affine invariant parallel SIFT feature extraction algorithm, parallel feature matching algorithm and improved parallel bundle adjustment algorithm.Firstly, In order to get radial distortion invariant SIFT algorithm, this paper analyzes the radial-distorted image camera model. This model is introduced into the SIFT Gaussian function, and it corrects the gradient of the feature descriptor. Affine transformation model is described in this paper, it simulates the change of camera optical axis to preprocess images to eliminate the image’s affine transformation. Simultaneously, the image pre-processing and parallel feature extraction algorithm have been obtained by GPU.Then, this paper analyzes the structure of feature matching algorithm:it firstly processes the image feature matching, and then removes bad matching feature points. When it processes the image feature matching, it firstly builds parallel kd-tree on GPU, then it uses the parallel priority method to search all the nearest and next nearest neighbor feature points. When it processes the removing bad matching points, it analyzes the RANSAC algorithm based on fundamental matrix, and improves this algorithm to fit the parallel processing. At last, this algorithm is achieved on GPU.Finally, this paper introduces the principle of bundle adjustment, and makes two improvements on the bundle adjustment algorithm:on the one hand, this paper puts forward weighted bundle adjustment algorithm, according to different uncertainty of the initial optimized parameters; On the other hand, this paper uses pre-conjugate gradient method to solve equations in bundle adjustment algorithm. At last, it achieves parallel bundle adjustment algorithm by using GPU.