Unsupervised multi-spectral image feature point matching under different lighting scenes
Junchong Huang, Wei Tian, Yongkun Wen, Zhan Chen, Yuyao Huang · 2021 5th CAA International Conference on Vehicular Control and Intelligence (CVCI) · 2021
In the case of low illumination condition or strong illumination variation, traditional visual SLAM often fails in feature point extraction or tracking. In order to improve the illumination robustness of SLAM front-end and facilitate localization in the same scene of day and night, we propose the UnsuperPoint-CLR network which integrates multi-spectral information, namely the RGB and thermal imaging, with an improved descriptor loss for image feature point matching. In this paper, two kinds of single mode information are evaluated in terms of repeatability score, localization error, homography estimation and matching score. An RGB-T fusion architecture based on RGB and thermal image features is established and trained by a self-supervised learning method, which learns the cross-modal feature matching. The proposed approach is trained and validated on our proposed RGB-T aligned dataset. The results reveal that the learned descriptor is with an improved illumination robustness compared with traditional methods.