Robust vision algorithms for quadruped soccer robots

This paper presents a robust vision system based on semi-automatic color calibration, color segmentation and image high-level analysis for quadruped soccer playing robots. The algorithms presented were used in the context of FC Portugal legged league team that participated since 2003 in RoboCup – Robotic Soccer World Championship. Controlled experiments with variable lightning conditions are analyzed in order to conclude the robustness of our vision system. pertise in RoboCup simulation league and on developing complex simulators [16, 17], we have built a very simple legged league simulator (with very simple models of the robots) enabling us to test different positioning strategies. Afterwards we have bought Sony ERS210A robots and moved our code from the simulator to the real robots. For that, we have used over the years, CMPack02 [23] and UNSW03 [7, 8, 22] codes as the base. We have applied over the base code, several previously researched methodologies developed and tested in our teams in other RoboCup leagues (Simulation, SmallSize, Middle-Size and Coach Leagues). From FC Portugal [10] (champion of RoboCup simulation league in 2000, European champion in 2000 and German Open Winner in 2001) we introduced simple versions of SBSP – Situation Based Strategic Positioning [17, 18], ADVCOM – Advanced Communications [11, 17] and DPRE – Dynamic Positioning and Role Exchange [17, 18]. From our FC Portugal Coach (Coach Competition champion in 2002), we have taken our tactical structure and coaching language [19]. From 5DPO teams [1] (small-size 3rd in RoboCup 1998, German Open Winners in 2001 and 2nd in 2002) we have taken the base vision system and most of our navigation algorithms [9, 15]. Our vision algorithms were then extended in order to enable robust color segmentation, automatic calibration and high-level image analysis. This paper describes briefly these extensions and the results achieved by our vision system in variable lightening conditions. The paper is organized as follows. Section 2 presents our vision module. Section 3 describes the localization module and the world state information used for highlevel decisions. Section 4 presents some results and section 5 the paper conclusions.