Performance analysis of algorithms for detecting roof faces in airborne laser scanner data

Over the past years, several algorithms have been proposed for extracting planar roof faces from airborne laser point clouds. These roof patches are used for automatic reconstruction of three dimensional building models. The models are useful in applications such as cadastral, urban planning, tourism, facility management, telecommunication network planning, real estate, vehicle and pedestrians’ navigation, and upcoming applications such as location based services and augmented reality. The intention of this research was to evaluate performance of different algorithms proposed for detecting roof faces. Analysis was done for performance in terms of completeness and accuracy of extracted faces, and for processing speed. The approaches to roof faces detection from laser data were reviewed and a selection made for two algorithms that were analysed. These are Plane growing of points using 3DHough Transform seed selection and Plane growing using TINmeshes. A working software was obtained for the former and literature transformed to a working software tool for the latter. Parameters of the algorithms were tuned to the optimum, used in experiments with data at different levels of point spacing to extract roof faces of varying complexity and then results analysed. Reference data sets were created interactively and a methodology, involving performance matrices, for comparing reference to detected faces was developed. A software tool was prepared for computing these matrices and then determining the number of correctly segmented roof faces, over segmented, under segmented, missed faces and noise segmentations. The tool also computes accuracy of the correctly detected faces in terms of angle and distance between the detected and corresponding faces in reference. All large faces can be detected by both algorithms. Faces with small objects such as chimneys and dormers are correctly detected by the 3D Hough algorithm while they are over segmented by the TIN algorithm. The geometric accuracy of the detected faces is high for both algorithms, a little higher for Hough than TIN. The performance for both algorithms gradually decreases with increase in point spacing. The speed of processing is in the tune of seconds for Hough while TIN takes longer up to hours depending on data size. The developed methodology can be used for other analyses while already, with the 3D Hough algorithm, roof faces can be reliably detected. Three dimensional building models can thus be more accurately and automatically constructed.