Detection and Robust Estimation of Cylinder Features in Point Clouds

The objective of this work is to develop new methods for efficient automatic 3D modeling of existing industrial installations from point cloud data. Traditionally, cylinder feature extraction algorithms utilize 5D Hough transforms, resulting in impractically high computational complexity. A more efficient approach uses a 2D Hough transform to estimate orientation followed by a 3D Hough transform to detect position, but still has extensive runtimes and lacks robustness in dense point cloud data. This work endeavors to (1) further decrease the runtime for cylinder feature extraction by implementing a coarse-to-fine approach, and (2) improve the robustness of the algorithm in detecting multiple cylinders by applying a clustering algorithm. In the coarse-to-fine approach, an initial estimate of the cylinder feature is quickly generated by coarsely sampling the Hough space. Subsequently, the search space is iteratively restricted based on the previous estimate while increasing sampling density to generate continually improving feature estimates until a stop criterion is reached. Results show that the implemented coarseto-fine approach yielded an improvement in orientation estimate accuracy of 20% while reducing runtime by 74%. To improve the robustness of the Hough transform in the presence of multiple cylinders, a clustering technique is implemented on the accumulator. First, cells in the accumulator with small number of accumulations are discarded to facilitate the computation of a hierarchical tree. Clustering is then applied to group the remaining cells into clusters representing different cylinders. This method improves robustness as well as accuracy of feature extraction in point cloud data with diverse cylinders.