Abstract
Camera tracking and geometric sensing in good conditions are considered solved problems by many. What many researchers ignore are more challenging applications, such as underwater and in dark and foggy environments [59]. This thesis highlights problems that are prevalent in these situations and presents ideas and improvements for enhancing the capabilities of camera tracking and geometric sensing methods in visually challenging applications. It explores new alternatives to classic methods for camera tracking by introducing robust image features for frame-to-frame pose estimation, enhancing them with machine learning algorithms to identify and ignore keypoints placed on marine snow particles. It covers new techniques in volumetric modeling to not only densely model the hard scene geometry but also implicitly account for the scattering medium. Further, using this as a map representation for accurate camera tracking within the scene. The thesis comprises five different papers and presents them as a compendium.
The first paper details the creation of a synthetic underwater dataset to act as a test bed for several of the following papers. The second paper investigates the detection of marine snow for Simultaneous Localization And Mapping (SLAM) approaches by focusing on detected keypoints and suppressing key points considered marine snow. Using keypoints allows a remarkably small and efficient classification model to attain accuracies in the high 90% more than enough for a RANSAC method [66] to remove the few remaining errors. The third paper combines the robustness of the features of geometric keypoints with the accuracy of photometric matching in a general attempt at increasing the performance of detection and matching methods. The paper concludes that no significant gain in accuracy above the projection error baseline could be reached. The fourth paper exploits the volumetric part of volumetric radiance fields to model volumetric media like fog and remove it from the model, resulting in an unobstructed view of the scene. We successfully removed the fog using this method, and the method’s effectiveness in removing very heavy fog surprised even us. The fifth and final paper presents a method for uniting the geometric modeling of volumetric radiance fields and camera tracking into a mathematically simple and computationally efficient method. Our method exceeded competing methods in modeling accuracy and camera tracking, given the same computation time.