Monocular Visual Odometry for Autonomous Underwater Navigation : An analysis of learning-based monocular visual odometry approaches in underwater scenarios

Abstract: Visual Odometry (VO) is the process of estimating the relative motion of a vehicle by using solely image data gathered from the camera. In underwater environments, VO becomes extremely challenging but valuable since ordinary sensors for on-road localization are usually unpractical in these hostile environments. For years, VO methods have been purely based on Computer Vision (CV) principles. However, the recent advances in Deep Learning (DL) have ushered in a new era for VO approaches. These novel methods have achieved impressive performance with state-of-the-art results on urban datasets. Nevertheless, little effort has been made to push learning-based research towards natural environments, such as underwater. Consequently, this work aims to bridge the research gap by evaluating the effectiveness of the learning-based approach in the navigation of Autonomous Underwater Vehicles (AUVs). We compare two learning-based methods with a traditional feature-based method on the Underwater Caves dataset, a very challenging dataset collected in the unstructured environment of an underwater cave complex. Extensive experiments are thus conducted training the models on this dataset. Moreover, we investigate different aspects and propose several improvements, such as sub-sampling the video clips to emphasize the camera motion between consecutive frames, or training exclusively on images with relevant content discarding those with dark borders and representing solely sandy bottoms. Finally, during the training, we also leverage underwater images from other datasets, hence acquired from different cameras. However, the best improvement is obtained by penalizing rotations around the x-axis of the camera coordinate system. The three methods are evaluated on test sequences that cover different lighting conditions. In the most favorable environments, although learning-based methods are not up to par with the feature-based method, the results show great potential. Furthermore, in extreme lighting conditions, where the feature-based baseline sharply fails to bootstrap, one of the two learning-based methods produces instead qualitatively good trajectory results, revealing the power of the learning-based approach in this peculiar context.