The acquisition of knowledge about the world is an essential endeavour of science. However, performing on-site observations at a global scale is unfeasible, therefore remote sensing is an appealing alternative. Over the last two decades, LiDAR (Light Detection And Ranging), an active remote sensing technique, has gained significant adoption. LiDAR allows acquiring a 3D record of the target scene in the form of point cloud with high accuracy. The goal of this thesis is to develop efficient methods for the classification of LiDAR data. For this, both general-purpose methods (segmentation and classification) and application-specific methods (building and road points extraction) are proposed, which have been efficiently implemented in a middle-to-low level language with an optimal spatial indexing and multi-core parallelisation. Experiments showed that the proposed growing criteria of the region-growing segmentation improved the segmentation global score from 1.053 a 1.019 (less is better) and the overall classification accuracy by 3.57% with nearly a 10% improvement in the building class accuracy. The rule-based classification achieved without training an overall accuracy of 83.3%, with a class accuracy of 88.4%, 86.9% and 70.5%, for ground, vegetation and buildings, respectively. The building point extraction yielded correctness, completeness and quality of 92.39%, 92.64% and 86.97%, respectively, with a speedup of 2.08 over using the general-purpose classifier for this task. The road point extraction was capable of estimating the intensity threshold and showed correctness, completeness and quality of 83%, 93% and 78%, respectively, in the most complete road dataset we are aware of. The methods for road delineation and characterisation demonstrated promising experimental results. Also, the feasibility of real-time processing is explored. A method for ground filtering exploiting the scan-line acquisition pattern of the LiDAR data showed satisfactory results under visual inspection, and achieved, in a synthetic reference data, Type I and Type II errors of 0.25% and 2.71%, respectively, and a kappa value of 88.59%. The method has a high computational performance, as it is capable of processing up to 1 million points per second in a workstation. The implementation was ported into a low-cost development board using FPGA acceleration, where experiments demonstrated that it is achievable to process up to 250,000 points per second in a higher-end reconfigurable system-on-chip, so it can cope with the data acquisition rates of the current lightweight scanners used in UAVs. Furthermore, a point cloud visualisation tool, namely OLIVIA, is presented. OLIVIA is an OpenGL-based open-source project implemented in Java, that offers an easy way to create customised visualisation and the capability of 3D stereoscopic view.
Keywords: Airborne LiDAR, Point Clouds, Segmentation, Classification, Extraction