Sky Segmentation by Fusing Clustering with Neural Networks Ali Pour Yazdanpanah 1 , Emma E. Regentova 1 , Ajay Kumar Mandava 1 , Touqeer Ahmad 2 , and George Bebis 2 (1) Dept. of Electrical and Computer Engineering, University of Nevada, Las Vegas 4505 S. Maryland Parkway Las Vegas, Nevada 89154-4026 (2) Dept. of Computer Science & Engineering, University of Nevada, Reno 1664 N. Virginia Street Reno, Nevada 89557-0208 Abstract. Sky segmentation is an important task for many applications related to obstacle detection and path planning for autonomous air and ground vehicles. In this paper, we present a method for the automated sky segmentation by fus- ing K-means clustering and Neural Network (NN) classifications. The perfor- mance of the method has been tested on images taken by two Hazcams (ie., Hazard Avoidance Cameras) on NASA’s Mars rover. Our experimental results show high accuracy in determining the sky area. The effect of various parame- ters is demonstrated using Receiver Operating Characteristic (ROC) curves. 1 Introduction NASA's Mars Exploration Rover mission (MER) and Mars Science Laborato- ry mission (MSL) are ongoing robotic space missions involving three rovers, explor- ing Mars. Two of the most important tasks during their missions are route planning, and path finding. The first step in route planning and path finding is to determine the suitability of the terrain for traversal. This includes extracting appropriate features for assessing rover navigation difficulty. To accomplish this task, accurate sky segmenta- tion is required. This is not an easy task, however, due to the diversity of skyline shapes (boundaries between sky regions and non-sky areas) and clutter like clouds. There are two main categories of sky segmentation found in computer vision litera- ture. In the first category, the problem is addressed as finding a horizon line/sky line which mostly depends on edge detection and some post processing on top of detected edges. The regions above the horizon line are labeled as sky whereas the regions below the horizon are labeled as non-sky. In the second category of sky segmentation, the problem is formulated as a pixel wise classification problem so every pixel in the given image gets a sky or non-sky label. In [4] Lie et al. have presented an edge based horizon line detection method. They formulate the horizon finding problem as a mul- ti-stage graph problem where a shortest path is found extending from the left most column to the right most column. A Sobel/Canny edge detector is applied on the giv- en gray scale image. The detected edges are used as graph nodes and links with zero or higher costs are placed between nodes if they are adjacent or have gaps. The gaps