Calibration of full-waveform airborne laser scanning data for 3D object segmentation

Abstract

Airborne Laser Scanning (ALS) is a fully commercial technology, which has seen rapid uptake from the photogrammetry and remote sensing community to classify surface features and enhance automatic object recognition and extraction processes. 3D object segmentation is considered as one of the major research topics in the field of laser scanning for feature recognition and object extraction applications. The demand for automatic segmentation has significantly increased with the emergence of full-waveform (FWF) ALS, which potentially offers an unlimited number of return echoes. FWF has shown potential to improve available segmentation and classification techniques through exploiting the additional physical observables which are provided alongside the standard geometric information. However, use of the FWF additional information is not recommended without prior radiometric calibration, taking into consideration all the parameters affecting the backscattered energy. The main focus of this research is to calibrate the additional information from FWF to develop the potential of point clouds for segmentation algorithms. Echo amplitude normalisation as a function of local incidence angle was identified as a particularly critical aspect, and a novel echo amplitude normalisation approach, termed the Robust Surface Normal (RSN) method, has been developed. Following the radar equation, a comprehensive radiometric calibration routine is introduced to account for all variables affecting the backscattered laser signal. Thereafter, a segmentation algorithm is developed, which utilises the raw 3D point clouds to estimate the normal for individual echoes based on the RSN method. The segmentation criterion is selected as the normal vector augmented by the calibrated backscatter signals. The developed segmentation routine aims to fully integrate FWF data to improve feature recognition and 3D object segmentation applications. The routine was tested over various feature types from two datasets with different properties to assess its potential. The results are compared to those delivered through utilizing only geometric information, without the additional FWF radiometric information, to assess performance over existing methods. The results approved the potential of the FWF additional observables to improve segmentation algorithms. The new approach was validated against manual segmentation results, revealing a successful automatic implementation and achieving an accuracy of 82%.