A Machine Learning Approach to Horizon Line Detection Using Local Features

Abstract

Planetary rover localization is a challenging problem since no conventional methods such as GPS, structural landmarks etc. are available. Horizon line is a promising visual cue which can be exploited for estimating the rover’s position and orientation during planetary missions. By matching the horizon line detected in 2D images captured by the rover with virtually generated horizon lines from 3D terrain models (e.g., Digital Elevation Maps(DEMs)), the localization problem can be solved in principle. In this paper, we propose a machine learning approach for horizon line detection using edge images and local features (i.e., SIFT). Given an image, first we apply Canny edge detection using various parameters and keep only those edges which survive over a wide range of thresholds. We refer to these edges as Maximally Stable Extremal Edges (MSEEs). Using ground truth information, we then train an Support Vector Machine (SVM) classifier to classify MSEE pixels into two categories: horizon and non-horizon. Each MSSE pixel is described using SIFT features which becomes input to the SVM classifier. Given a novel image, we use the same procedure to extract MSSEs; then, we classify each MSEE pixel as horizon or non-horizon using the SVM classifier. MSEE pixels classified as horizon are then provided to a Dynamic Programming shortest path finding algorithm which returns a consistent horizon line. In general, Dynamic Programming returns different solutions (i.e., due to gaps) when searching for the optimum horizon line in a left-to-right or right-to-left fashion. We use the actual output of the SVM classifier to resolve ambiguities in places where the left-to-right and right-to-left solutions are different. The final solution, is typically a combination of edge segments from the left-to-right or right-to-left solutions. Moreover, we use the SVM classifier to fill in small gaps in the horizon line; this is in contrast to the traditional dynamic programming approach which relies on mere interpolation. We report promising experimental results using a set of real images.