Comparison of Local Visual Feature Detectors and Descriptors for the Registration of 3D Building Scenes

Abstract

Three-dimensional (3D) as-built geometric models are useful for many building assessment and management tasks. However, the current process of creating such models is labor-intensive. A significant amount of manual work is required to register the remote sensing data captured from multiple scans into one scene (i.e., scene registration). To automate the registration work, several research studies have been developed to automate the registration process by the detection and matching of common visual features in consecutive scans. This paper investigates the effectiveness of different combinations of common visual feature detectors and descriptors that have been widely used in the scene registration of 3D buildings. The evaluation criteria include registration accuracy and speed. The feature detectors and descriptors have been tested in a total of 31 realistic building scenarios. The results show that the combination of the scale-invariant feature transform feature detector and descriptor reached more accurate results than the others. The fastest speed is achieved by the use of an oriented binary robust independent elementary features (ORB) detector in combination with the speeded-up robust features or ORB descriptor.