Research On Key Techniques Of 3D Surface Reconstruction Based On Depth Camera

The three-dimensional (3D) surface reconstruction technology mainly is referred to restore the 3D information of the target objects from the image or the image sequence, which has broad application foreground and business value in the field of telemedicine, immersive virtual interaction, preservation of cultural relics and 3D printing. As the key technology and research focus in the field of computer vision, augmented reality and human-computer interaction,3D surface reconstruction is one of the major challenges in basic research and applied research. With the development of the depth cameras,3D surface reconstruction based on depth cameras becomes a new hotspot in recent years. Our research focuses on improving the precision and scale of 3D surface reconstruction, including the issues of pretreatment of the depth data, registration of the range image, point cloud data fusion and depth camera relocation. The major innovations and contributions of this paper are describes as follows:1. This paper studies the optical imaging system of ToF depth camera and establishes the model of ToF camera noise. An innovative filtering algorithm based on weighted least squares framework is proposed for ToF depth map with the noise model. Compared with the common range image filters, the proposed algorithm can not only achieve better performance in edge preservation but also eliminate the noise of ToF depth map more effectively.2. Due to the restriction of computer's memory capacity and shooting angle of depth cameras, the region of 3D surface reconstruction is always be limited to small-scale indoor scenes. In order to achieve large-scale scene reconstruction under existing hardware conditions, this paper designs an effective active volume shift strategy to extend the tracking range without increment on storage resource. Furthermore, this paper proposes three local optimization methods including motion predicting, weighted ICP and multi-model frame combined estimation to improve the tracking stability and decrease the accumulative error. In KinectFusion, the scene is represented as an uniform grid of voxels. Though the uniform grid of voxels is suitable for parallel computation in graphics hardware, most of the storage is wasted, since the geometry is very sparse in the scene volume. In order to reduce the memory cost and save the computation time, this paper develops an algorithm called "octrees forest" which combines both volumetric model and octree method for an adaptive grid representation.3. There is a problem of failure of camera tracking in the process of 3D surface reconstruction. To solve this problem, this paper adopts an approach for camera relocalization which employs a random forest that is capable of inferring an estimate of each pixel's correspondence to 3D points in the scene's world coordinate frame. The camera pose is inferred using a single acquired image when the camera tracking is failed. Compared to the approaches based on image matching, our method uses only simple depth and RGB pixel comparison features, and does not require the computation of feature descriptors.