An Experimental Comparison of Image Feature Detectors and Descriptors applied to Grid Map Matching J.L. Blanco, J. Gonzalez, J.A. Fern´ andez-Madrigal * Technical Report April 28, 2010 Abstract Applying computer vision feature detectors and descriptors to occu- pancy grids has important practical applications for the problem of grid map matching in mobile robot localization and mapping, although this approach has received little attention by the community. This review presents a thorough performance evaluation for several combinations of detectors (Harris, KLT, SIFT and SURF) and descriptors (SIFT, SURF and circular patches) using maps obtained from real datasets. It is shown how a combination of the Harris or KLT detector with circular patch descriptors provides the best results in both computation time and clas- sification success ratio. Keywords – Occupancy grid maps, Image registration. 1 Introduction Occupancy grid maps are one of the most widespread world representations employed in the mobile robot community [10, 11]. Most mainstream approaches to Simultaneous Localization and Mapping (SLAM) that rely on grid maps need to perform certain operations on these grids, namely: update them from sensory data and estimate the sensor observation likelihood [21]. Matching pairs of grid maps is also required when dealing with hybrid metric-topological models [3, 5, 8] or multi-robot mapping [2]. Several works have focused on problems such as building grids from raw sen- sory data [7, 20] and the more complex task of estimating observation likelihoods ([4, 18, 21]). In contrast, grid matching has received relatively little attention * Authors are with the Department of System Engineering and Automation, University of M´ alaga, Spain. 1