Robust and Efficient Object Recognition for a Humanoid Soccer Robot Alexander H¨ artl 1 , Ubbo Visser 1 , and Thomas R¨ ofer 2 1 University of Miami, Department of Computer Science, 1365 Memorial Drive, Coral Gables, FL, 33146 USA {a.haertl,visser}@cs.miami.edu 2 Deutsches Forschungszentrum f¨ ur K¨ unstliche Intelligenz, Cyber-Physical Systems, Enrique-Schmidt-Str. 5, 28359 Bremen, Germany thomas.roefer@dfki.de Abstract. Static color classification as a first processing step of an ob- ject recognition system is still the de facto standard in the RoboCup Standard Platform League (SPL). Despite its efficiency, this approach lacks robustness with regard to changing illumination. We propose a new object recognition system where objects are found based on color similarities. Our experiments with line, goal, and ball recognition show that the new system is real-time capable on a contemporary NAO (ver- sion 3.2 and above). We show that the detection rate is comparable to color-table-based object recognition under static lighting conditions and substantially better under changing illumination. 1 Introduction Color-based recognition of geometrically simple objects is a well known problem that has been extensively studied for over two decades [18,5]. Prominent exam- ples for successful image processing methods are edge detection [2,10], region- growth algorithms [21,8], and histogram-based algorithms [12,19,14]. A popular approach is based on edge detection and subsequent Hough transformation [4] in which the authors show a method for line detection that can be used for more general curve fitting. Although the literature shows a broad range of varia- tions and implementations of the mentioned approaches, only a few can be used for embedded systems that are constrained by limited resources such as time, memory, and/or CPU power. The RoboCup Soccer environment demands efficient real-time object recogni- tion. Many systems are still based on fixed color tables that are similar or based on the CMVision system [1]. It is well suited for static lighting conditions but lacks robustness when illumination varies. R¨ofer [15] improved the robustness by introducing ambiguous color classes and delaying hard decisions to a later pro- cessing stage. Reinhardt [14] uses different heuristics applied to color histograms to cope with variations in illumination. We propose a new object recognition system where objects are found based on color similarities. As a first step, a subsampling is created considering the S. Behnke et al. (Eds.): RoboCup 2013, LNAI 8371, pp. 396–407, 2014. c  Springer-Verlag Berlin Heidelberg 2014