Navigating Beyond GPS: Horizon Detection for Autonomous Navigation

In a scenario where a GPS signal is unavailable, navigation becomes more challenging. To enable precise navigation, visual features of the surroundings can be interpreted to direct vessels. We propose to use the visible horizon line as a novel solution. The visible horizon line is constructed from features at the intersection between the sky, land masses, and the waterline. Accurate extraction relies on a robust model operating in adverse weather and equipment conditions. We propose to tackle this new problem with an approach based on semantic boundary prediction with a framework that primarily looks at deep supervision for feature conditioning and differential edge detection to incorporate strong priors early in the training process. We also use semantic segmentation as a support task to provide a strong supervision signal and features to the semantic boundaries prediction task. We conduct experiments on the dataset to evaluate the model in various scenarios. Our results show that our framework predicts high-quality boundaries and segmentation masks over various datasets and domains with the ability to perform in low-data scenarios. We demonstrate that the combination of the labelling and the edge priors improve over the baseline by 32-44% on the semantic boundary estimation task.