Automatic Illumination‐Invariant Image‐to‐Geometry Registration in Outdoor Environments

Abstract

Image-to-geometry registration is the basis of many applications for texturing and interpreting 3D surface models. Feature-based matching is an established, automatic approach which creates 2D-3D correspondences based on salient points and their radiometric neighbourhood. This paper presents an experimental approach for assessing the accuracy of several matching algorithms in challenging imaging environments that are subject to significant outdoor illumination variations. Furthermore, a collection of accuracy assessment metrics and quality heuristics emerge from the presented approach to guide a user during the examination of registration results. As a result of the experiments, two novel salient point descriptor matching combinations outperform the standard scaleinvariant feature transform (SIFT) operator on the task of image-to-image and image-to-geometry registration under varying illumination conditions.The Photogrammetric Record

Photogrammetric water level determination using smartphone technology

“…Kehl et al. () pointed out that radiometric variances, such as irregular illumination caused by casting shadows or overexposure, as well as environmental conditions like vegetation cover, are the principal reasons for SIFT‐based matching failures. Regarding the case study given in Fig.…”

“…() or Kehl et al. (). It has been previously demonstrated that image registration can work just by using terrestrial laser scanning (TLS) and its return signal intensity together with an RGB camera image (Wang et al., ; Guislain et al., ).…”

“…However, this requires very stable radiometric (illumination) conditions and a prohibitively complex mutual information (MI) framework to retrieve reliable 2D–3D point correspondences to further determine the exterior orientation. Registering outdoor photographs with significant radiometric variance requires colour information for the photograph as well as the 3D point cloud of which the synthetic image is rendered (Kehl et al., ).…”

“…Sattler et al (2011) and Sibbing et al (2013) have provided examples for this approach, which has also been implemented in this paper. In the case of slight illumination variations, local contrast adjustments can be applied before performing feature detection and description, for example, by applying Wallis or Gaussian filtering as described in or Kehl et al (2017). It has been previously demonstrated that image registration can work just by using terrestrial laser scanning (TLS) and its return signal intensity together with an RGB camera image Guislain et al, 2016).…”

Nationwide Point Cloud—The Future Topographic Core Data

“…While radiometric calibration of ALS can be performed to some extent even with natural targets [98], the radiometric calibration of TLS (and MLS) is more challenging, both due to instrumental effects and highly altering environments and ranges encountered [99]. Further, if imaging sensors are used to produce spectral information, their performance limitations affect the final result [100]. While multispectral systems are already available for ALS, they are still absent from TLS and MLS.…”