Research On 3D Reconstruction Technology With Depth Cameras

Reconstructing 3D models from the data captured by the depth camera is an important subject in computer vision.It is widely used in areas such as medical imaging,games and multimedia,human-computer interaction,augmented reality because of easy data acquisition,low cost and flexibility.At present,the depth camera based 3D reconstruction method is relatively mature.However,there are two key problems in the field: 1)the rotation of camera,illumination and occlusion problems in the process of scene data acquisition will affect the pose estimation of the camera,and then affect the reconstruction accuracy;2)low power depth cameras have the disadvantages of low resolution and large noise,which makes it difficult to construct three-dimensional models with rich texture information.Based on the problems of 3D reconstruction mentioned above,this thesis develops related research,the main research contents and innovations are as follows:1.In the aspect of deep camera tracking,this thesis makes full use of the co-planarity constraints,geometric relationship constraints and feature correspondence constraints between different objects and planar regions in the image,and proposes a fine-to-coarse camera tracking algorithm.The increasing window size results in a globally consistent camera pose results,which effectively improves the accuracy and robustness of the pose estimation in different scenarios.2.The depth camera can capture multiple images in a limited time,which leads to the redundancy of the image sequence.In the process of hand-held photography,motion blur will occur due to excessive movement or trembling.In this thesis,a keyframe selection algorithm based on discrete wavelet transform is proposed.The image is decomposed into multiple highfrequency subbands,the weighted sum of the logarithmic energies of these subbands is calculated,and the high-quality images are selected from the image sequence to enhance the reconstruction result.3.In the optimization process of 3D geometric model,the proposed algorithm uses the spherical harmonic function to construct the global lighting model.To obtain a high-accuracy model with rich texture,lighting coefficients and the sparse 3D model are jointly utilized to implement the model refinement and texture mapping.Through extensive experiments and evaluations,the results show that the keyframe selection algorithm proposed in the thesis can extract high-quality images from redundant input image sequences.Camera tracking algorithm and optimization algorithm can effectively calculate the camera pose,improve the details of the reconstruction model,and thus obtain a globally consistent scene geometry from the input of the depth camera.