Vision-based Navigation Solution for Autonomous Underwater Vehicles

Vehicle navigation is fully autonomous when the system is capable of planing its path and execute it without human intervention. This research aims at introducing an AI-based approach for visual navigation in underwater environments. To achieve this, several challenges have to be overtaken, such as segmenting the images to filter the floating clutter typical in underwater environments. First, an annotated dataset with pairs of input images and segmentation grounds truths is essential for training a state-of-the-art AI model. Second, choosing a model adequate for image segmentation and training it. Finally, evaluate if this methodology improves the accuracy of visual navigation and scene reconstruction algorithms, such as online and offline SLAM. This approach achieved state-of-the-art results on the segmentation task, with 93% pixel accuracy and 85% IoU. At last, it was concluded that using the segmentation masks produced by the fully convolutional network improves the results of using offline and online SLAM algorithms.