The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B1-2022
https://doi.org/10.5194/isprs-archives-XLIII-B1-2022-157-2022
https://doi.org/10.5194/isprs-archives-XLIII-B1-2022-157-2022
30 May 2022
 | 30 May 2022

JOINT BUNDLE ADJUSTMENT OF THERMAL INFRA-RED AND OPTICAL IMAGES BASED ON MULTIMODAL MATCHING

A. Sledz and C. Heipke

Keywords: thermal infrared and optical image registration, multimodal feature matching, phase congruency, bundle adjustment

Abstract. Despite the fact that Thermal Infrared (TIR) cameras have been in use for decades, processing of TIR images poses a variety of challenges when compared to optical images, which are captured in the visible part of the electromagnetic spectrum. The estimation of the exterior orientation of TIR cameras by bundle adjustment is a difficult task due to the limited geometric resolution of a TIR camera and the low image quality in terms of contrast and texture compared to optical images. Optical images have a potential to increase TIR external orientation accuracy by incorporating them into a joint bundle adjustment. However, because the modality gap between those two image types is large, classical point matching algorithms typically fail to find matches, making processing both image types in the joint bundle challenging. In order to locate matching points in both modalities, this study suggests using the Edge Histogram Descriptor (EHD) in the frequency domain representation of the images based on phase congruency. To properly allocate edges from the phase congruency, which are then employed in EHD, non-maximum suppression and hysteresis thresholding are used. Considering that both sensors are fixed rigidly to a single platform, the search region for the matching point candidate of the TIR image is determined based on stereo calibration of a thermal/optical stereo setup combined with geometric constraints. The final matching is based on the cosine distance, while RANdom SAmple Consensus (RANSAC) is used in order to eliminate outliers. The findings of this study show that using a joint bundle adjustment with optical images versus a bundle adjustment only with TIR images improves TIR image orientation, which is supported by the increased accuracy of the adjusted Ground Control Point (GCP) coordinates.