Horizon detection and tracking in sea-ice conditions using machine vision

An automated process is proposed for horizon detection and tracking using machine vision cameras and in polar, sea-ice conditions. These conditions present unique challenges for machine vision applications, such as a large amount of clutter (e.g. icebergs) and secondary edge lines from broken ice pieces. The process is divided in two parts: a more computationally expensive, yet robust detection algorithm in the first stage, based on Convolutional Neural Networks, and used to detect the horizon line in an arbitrary sea-ice image; followed by a tracking algorithm, responsible of efficiently detecting the horizon line in the subsequent images of a sequence. We propose two tracking algorithms, one based on the traditional Canny and Hough line detection methods; and a second novel approach using entropy as a measure of randomness, to segment between sea-ice and sky. Our automated process was compared to manually obtained ground-truth data and the results indicate good agreement, especially for the texture-based tracking algorithm.