TOFENetTopographic Features Extraction Network

Topographic features found in ground-based natural images contain information that is useful for a variety of applications including locationestimation and navigation. Traditionally these features have been manually labeled by analysts which is costly and time consuming, especiallyconsidering the volume of readily available data. We propose a novel method for extracting topographic features from still images and videousing a state-of-the-art deep learning based approach. Cues gathered from RGB images are used to locate topographic features in individualimages and include information about the relative size, position, and context of each feature. Furthermore, by tracking these features withinframes of a video the system can reliably mitigate false positives and gather a more comprehensive understanding of the environment beingimaged. Using carefully chosen training datasets our approach learns to estimate features such that the output is robust to changes in lighting,weather and environment. The proposed system will provide a pixel-wise labelling of topographical features, along with useful contextualmetadata about each feature.