| Image feature extraction and matching are crucial techniques in computer vision that have been extensively developed over the years and used in various scenarios with remarkable results.Nonetheless,the current image matching algorithms encounter significant difficulties when dealing with extreme environments,such as lunar surfaces and low-textured environments,due to challenges such as wide baseline,poor texture,and illumination variations.In this thesis,we concentrate on the development of image matching algorithms specially designed for lunar images with wide baselines and weak textures.We then integrate a robust feature extraction algorithm into the visual SLAM,enhancing its robustness in weak-textured scenes.This thesis proposes a self-supervised monocular wide-baseline lunar image matching algorithm,H-Warp-MDNet,to improve the low success rate of image feature extraction and matching algorithms.The proposed algorithm is based on plane homographic view synthesis and introduces new solutions in the image preprocessing,feature extraction network,and outlier rejection stages.To address the challenges posed by wide baseline images,homographic view synthesis is employed with prior inertial measurements between sites.The feature extraction network is trained in a self-supervised manner and the plane degeneration is considered in the outlier rejection stage.Experiments conducted on a real lunar image dataset demonstrate that the H-Warp-MDNet algorithm has a better performance in terms of matching success rate and accuracy when compared to current mainstream image matching algorithms.To enhance the success rate of matching lunar images with a wide baseline and reduce the manual labeling burden,we present a global attention lunar image matching algorithm based on depth view synthesis with stereo images,called Depth Warp-Lo FTR.We use sparse feature matching methods to generate sparse pseudo-ground-truth disparities for the rectified stereo lunar images at the same site.Then we finetune a stereo matching network with these disparities and perform3 D reconstruction for the lunar images at the same site.The inertial measurements between different sites are used to convert the original image into a new synthetic view for matching based on the depth of scenes,addressing the problem of low overlap and large viewpoint changes between images of different sites.In matching stage,we adopt a Transformer-based image matching network to improve matching performance in scenes with weak texture.Our algorithm is evaluated on a real lunar image dataset and achieves improved matching precision and success rates,thereby paving the way for visual localization of lunar vehicles during long-distance travel.To address the problem of tracking failure for visual SLAM in low-textured scenes,we propose a novel visual SLAM system called RWT-SLAM for highly weak-textured environments.We use existing matching networks to generate dense descriptors for images with poor texture,and employ a soft detection module for keypoint detection and high saliency descriptor selection.We trained a suitable dictionary for the bag-of-words model and implemented the design of the matching strategy and integration with the SLAM system.Experiments on public datasets such as TUM RGB-D and Open LORIS show that RWT-SLAM can effectively perform localization and mapping in highly weak-textured environments while maintaining high accuracy. |
