Planar Feature Extraction in Terrestrial Laser Scans Using Gradient Based Range Image Segmentation

The paper presents a new segmentation algorithm, to be applied to terrestrial lasers scans of urban environments. The algorithm works directly in a range image. This is the fundamental laser scan data structure as a laser recording can be regarded as a 2dimensional grid of range measurements. The horizontal and vertical axes of the grid denote the horizontal and vertical angles at which the scanner emits the laser beam, receives the reflections, and measures the distance (the range) between the instrument and the reflecting surface at those angles. The presented algorithm estimates for each measurement (pixel) in the range image three parameters of the 3D plane that contains the pixel: two angles (horizontal and vertical) and the distance between the plane and the origin. The estimates are based on the scan angles (the horizontal and vertical angles at which the laser beam was emitted from the scanner) and the image gradients, i.e. the rate of change in the distance that is observed between adjacent measurements. Since the three estimated parameters characterize a plane in 3D space, regions of adjacent pixels with similar parameter values are likely to be part of the same plane. Such pixels are grouped into segments by a region-growing image segmentation step, which takes the three parameters into account simultaneously. The overall algorithm uses two complementary strategies to deal with the measurement noise affecting the gradients, during the gradient calculation and the region growing steps respectively.