Pose estimation for robotic disassembly using RANSAC with line features

In this thesis, a new technique to recognize and estimate the pose of a given 3-D object from a single real image provided known prior knowledge of its approximate structure is proposed. Metrics to evaluate the correctness of a calculated pose are presented and analyzed. The traditional and the more recent approaches used in solving this problem are explored and the various methodologies adopted are discussed.;The first step in disassembling a given assembly from its image is to recognize the attitude and translation of each of its constituent components---a fundamental problem which is being addressed in this work. The proposed algorithm does not depend on uniquely identifiable 3D model surface features for its operation---this makes it ideally suited for object recognition for assemblies. The algorithm works well even for low-resolution occluded object images taken under variable illumination conditions and heavy shadows and performs markedly better when these factors are removed.;The algorithm uses a combination of various computer vision concepts such as segmentation, corner detection and camera calibration, and subsequently adopts a line-based object pose estimation technique (originally based on the RANSAC algorithm) to settle on the best pose estimate. The novelty of the proposed technique lies in the specific way in which the poses are evaluated in the RANSAC-like algorithm. In particular, line-based pose evaluation is adopted where the line chamfer image is used to evaluate the error distance between the projected model line and the image edges. The correctness of the computed pose is determined based on the number of line matches computed using this error distance. As opposed to the RANSAC algorithm where the search process is pseudo-random, we do an exhaustive pose search instead. Techniques to reduce the search space by a large amount are discussed and implemented.;The algorithm was used to estimate the pose of 28 objects in 22 images, where some images contain multiple objects. The algorithm has been found to work with a 3-D mismatch error of less than 2.5cm in 90% of the cases and less than 1cm error in 53% of the cases in the dataset used.