Horizon Line Detection: Edge-less and Edge-based Methods

Abstract

Planetary rover localization is a challenging problem due to unavailability ofconventional localization cues e.g. GPS, architectural landmarks etc. Hori-zon line (boundary segmenting sky and non-sky regions) nds its applicationsfor smooth navigation of UAVs/MAVs, visual geo-localization of mountain-ous images, port security and ship detection and has proven to be a promisingvisual cue for outdoor robot/vehicle localization.Prominent methods for horizon line detection are based on faulty as-sumptions and rely on mere edge detection which is inherently a non-stableapproach due to parameter choices. We investigate the use of supervisedmachine learning for horizon line detection. Specically we propose two dif-ferent machine learning based methods; one relying on edge detection andclassication while other solely based on classication. Given a query image;an edge or classication map is rst built and converted into a multi-stagegraph problem. Dynamic programming is then used to nd a shortest pathwhich conforms to the detected horizon line in the given image. For the rstmethod we provide a detailed quantitative analysis for various texture fea-tures (SIFT, LBP, HOG and their combinations) used to train an SupportVector Machine (SVM) classier and dierent choices (binary edges, classi-ed edge score, gradient score and their combinations) for the nodal costsfor Dynamic Programming. For the second method we investigate the use ofdense classication maps for horizon line detection. We use Support VectorMachines (SVMs) and Convolutional Neural Networks (CNNs) as our classi-er choices and use raw intensity patches as features. Dynamic Programmingis then applied on the resultant dense classier score image to nd the hori-zon line. Both proposed formulations are compared with a prominent edgebased method on three dierent data sets: City (Reno) Skyline, Basalt Hillsand Web data sets and outperform the previous method by a high margin.