Dense RGB-D SLAM and object localisation for robotics and industrial applications

Lan, Feiying ORCID: 0000-0002-8916-3134 (2023). Dense RGB-D SLAM and object localisation for robotics and industrial applications. University of Birmingham. Ph.D.

Text - Accepted Version
Available under License All rights reserved.

Download (72MB) | Preview


Dense reconstruction and object localisation are two critical steps in robotic and industrial applications. The former entails a joint estimation of camera egomotion and the structure of the surrounding environment, also known as Simultaneous Localisation and Mapping (SLAM), and the latter aims to locate the object in the reconstructed scenes. This thesis addresses the challenges of dense SLAM with RGB-D cameras and object localisation towards robotic and industrial applications.

Camera drift is an essential issue in camera egomotion estimation. Due to the accumulated error in camera pose estimation, the estimated camera trajectory is inaccurate, and the reconstruction of the environment is inconsistent. This thesis analyses camera drift in SLAM under the probabilistic inference framework and proposes an online map fusion strategy with standard deviation estimation based on frame-to-model camera tracking. The camera pose is estimated by aligning the input image with the global map model, and the global map merges the information in the images by weighted fusion with standard deviation modelling. In addition, a pre-screening step is applied before map fusion to preclude the adverse effect of accumulated errors and noises on camera egomotion estimation. Experimental results indicated that the proposed method mitigates camera drift and improves the global consistency of camera trajectories.

Another critical challenge for dense RGB-D SLAM in industrial scenarios is to handle mechanical and plastic components that usually have reflective and shiny surfaces. Photometric alignment in frame-to-model camera tracking tends to fail on such objects due to the inconsistency in intensity patterns of the images and the global map model. This thesis addresses this problem and proposes RSO-SLAM, namely a SLAM approach to reflective and shiny object reconstruction. RSO-SLAM adopts frame-to-model camera tracking and combines local photometric alignment and global geometric registration. This study revealed the effectiveness and excellent performance of the proposed RSO-SLAM on both plastic and metallic objects. In addition, a case study involving the cover of a electric vehicle battery with metallic surface demonstrated the superior performance of the RSO-SLAM approach in the reconstruction of a common industrial product.

With the reconstructed point cloud model of the object, the problem of object localisation is tackled as point cloud registration in the thesis. Iterative Closest Point (ICP) is arguably the best-known method for point cloud registration, but it is susceptible to sub-optimal convergence due to the multimodal solution space. This thesis proposes the Bees Algorithm (BA) enhanced with the Singular Value Decomposition (SVD) procedure for point cloud registration. SVD accelerates the speed of the local search of the BA, helping the algorithm to rapidly identify the local optima. It also enhances the precision of the obtained solutions. At the same time, the global outlook of the BA ensures adequate exploration of the whole solution space. Experimental results demonstrated the remarkable performance of the SVD-enhanced BA in terms of consistency and precision. Additional tests on noisy datasets demonstrated the robustness of the proposed procedure to imprecision in the models.

Type of Work: Thesis (Doctorates > Ph.D.)
Award Type: Doctorates > Ph.D.
Licence: All rights reserved
College/Faculty: Colleges (2008 onwards) > College of Engineering & Physical Sciences
School or Department: School of Engineering, Department of Mechanical Engineering
Funders: None/not applicable
Subjects: T Technology > TJ Mechanical engineering and machinery


Request a Correction Request a Correction
View Item View Item


Downloads per month over past year