Humanoid Localization on Robocup Field using Corner Intersection and Geometric Distance Estimation

Abstract

In the humanoid competition field, identifying landmarks for localizing robots in a dynamic environment is of crucial importance. By convention, state-of-the-art humanoid vision systems rely on poles located outside the middle of the field as an indicator for generating landmarks. However, in compliance with the recent rules of Robocup, the middle pole has been discarded to deliberately provide less prior information for the humanoid vision system to strategize its winning tactics on the field. Previous localization method used middle poles as a landmark. Therefore, robot localization tasks should apply accurate corner and distance detection simultaneously to locate the positions of goalposts. State-of-the-art corner detection algorithms such as the Harris corner and mean projection transformation are excessively sensitive to image noise and suffer from high processing times. Moreover, despite their prevalence in robot motor log and fish-eye lens calibration for humanoid localization, current distance estimation techniques nonetheless remain highly dependent on multiple poles as vision landmarks, apart from being prone to huge localization errors. Thus, we propose a novel localization method consisting of a proposed corner extraction algorithm, namely, the contour intersection algorithm (CIA), and a distance estimation algorithm, namely, analytic geometric estimation (AGE), for efficiently identifying salient goalposts. At first, the proposed CIA algorithm, which is based on linear contour intersection using a projection matrix, is utilized to extract corners of a goalpost after performing an adaptive binarization process. Then, these extracted corner features are fed into our proposed AGE algorithm to estimate the real-word distance using analytic geometry methods. As a result, the proposed localization vision system and the state-of-the-art method obtained approximately 3-4 and 7-23 centimeter estimation errors, respectively. This demonstrates the capability of the proposed localization algorithm to outperform other methods, which renders it more effective in indoor task localization for further actions such as attack or defense strategies.