Mobile Mapping of Transport Corridors and the Extraction of Assets from Video and Range Data
Laser scanning has spawned a renewed interest in automatic robust feature extraction. Three dimensional point cloud data obtained from laser scanner based mobile mapping systems commonly contain outliers and/or noise. The presence of outliers and noise means that most of the frequently used methods for point cloud processing and feature extraction produce inaccurate and unreliable results i.e. are termed non-robust. Dealing with the problems of outliers and noise for automatic robust feature extraction in mobile laser scanning 3D point cloud data has been the subject of this research.
his thesis develops algorithms for statistically robust planar surface fitting based on robust and/or diagnostic statistical approaches. The algorithms outperform classical methods such as least squares and principal component analysis and show distinct advantages over current robust methods including RANSAC and its derivations in terms of computational speed, sensitivity to the percentage of outliers or noise, number of points in the data and surface thickness. Two highly robust outlier detection algorithms have been developed for accurate and robust estimation of local saliency features such as normal and curvature. Results for articial and real 3D point cloud data experiments show that the methods have advantages over other existing popular techniques in that they (i) are computationally simpler, (ii) can successfully identify high percentages of uniform and clustered outliers, (iii) are more accurate, robust and faster than existing robust and diagnostic methods developed in disciplines including computer vision (RANSAC), machine learning (uLSIF) and data mining (LOF), and (iv) have the ability to denoise point cloud data. Robust segmentation algorithms have been developed for multiple planar and/or non-planar complex surfaces e.g. long cylindrical and approximately cylindrical surfaces (poles), lamps and sign posts extraction. A region growing approach has been developed for segmentation algorithms and the results demonstrate that the proposed methods reduce segmentation errors and provide more robust feature extraction. The developed methods are promising for surface edge detection, surface reconstruction and fitting, sharp feature preservation, covariance statistics based point cloud processing and registration. An algorithm has also been introduced for merging several sliced segments to allow large volumes of laser scanned data to be processed seamlessly. In addition, the thesis presents a robust ground surface extraction method that has the potential for being used as a pre-processing step for large point cloud data processing tasks such as segmentation, feature extraction, classification of surface points, object detection and modelling. Identifying and removing the ground then allows more efficiency in the segmentation of above ground objects.