Performance Evaluation of ICP, CPD and SVR Methods for Automatic Registration of Roof Point Clouds Extracted from Airborne LiDAR Data

Abstract

From the 2000s onwards, there has been a massive acquisition of LiDAR (Light Detection and Ranging) data in urban areas, what enabled several studies and applications in the most diverse fields, with increasingly available historical databases. As a consequence, a search for robust methods to manipulate these data emerges. The data registration methods are crucial for using multisource laser data both in terms of acquisition date and sensor. This article evaluates the performance of three registration methods: Iterative Closest Point (ICP), Coherent Point Drift (CPD) and Support Vector Registration (SVR). The methodology copes with the pre-processing of the LiDAR data to extract roofs of three different buildings, located on the Federal University of Paraná (UFPR) campus, in Curitiba city. The data were collected by the Optech ALTM Pegasus HD 500 sensor, with a frequency of 300 kHz, flight height of 1,600 m, average density of 1.71 points per m² and an IFOV of 25°. Finally, the registrations based on the three employed methods were obtained, of which their accuracies and processing times were assessed. The results showed that the CPD and SVR methods are great alternatives to overcome the limitations of the ICP, but particular emphasis should be laid on the performance of CPD and on the computational efficiency of SVR, which is suitable for handling noisy data.