From images to augmented 3D models: improved visual SLAM and augmented point cloud modeling
MetadataShow full item record
This thesis investigates into the problem of using monocular image sequences to generate augmented models. The problem is decomposed to two subproblems: monocular visual simultaneously localization and mapping (VSLAM), and the point cloud data modeling. Accordingly, the thesis comprises two major parts. The First part, including Chapters 2, 3 and 4, aims to leverage the system observability theories to improve the VSLAM accuracy. In Chapter 2, a piece-wise linear system is developed to model VSLAM, and two necessary conditions are proved to make the VSLAM completely observable. Based on the First condition, an instantaneous condition for complete observability, the "Optimally Observable and Minimal Cardinality (OOMC) VSLAM" is presented in Chapter 3. The OOMC algorithm selects the feature subset of minimal required cardinality to form the strongest observable VSLAM subsystem. The select feature subset is further used to improve the data association in VSLAM. Based on the second condition, a temporal condition for complete observability, the "Good Features (GF) to Track for VSLAM" is presented in Chapter 4. The GF algorithm ranks the individual features according to their contributions to system observability. Benchmarking experiments of both OOMC and GF algorithms demonstrate improvements in VSLAM performance. The second part, including Chapters 5 and 6, aims to solve the PCD modeling problem in a geometry-driven manner. Chapter 5 presents an algorithm to model PCDs with planar patches via a sparsity-inducing optimization. Chapter 6 extends the PCD modeling to quadratic surface primitives based models. A method is further developed to retrieve the high-level semantic information of the model components. Evaluation on the PCDs generated from VSLAM demonstrates the effectiveness of these geometry-driven PCD modeling approaches.