Visual color and field line recognition and exploitation for the RoboCup Standard Platform League

Abstract

In order to complete complex tasks, both humans and robots are bound by one of the most important senses, the visual sense, which helps them perceive the state of the environment that surrounds them. Robotic soccer, known as RoboCup, represents a complex, stochastic, real-time, multi-agent, competitive domain for autonomous robots. In such domains, the ability to understand the environment is critical for the accomplishment of the assigned task and is required for a range of activities, such as locomotion, coordination, and decision making. This thesis focuses on two specific aspects of robot visual perception, color and field line recognition, in order to complement prior work on this problem by our RoboCup team Kouretes. In RoboCup, the objects of interest are characterized by unique colors (orange ball, green field, yellow goalposts, white lines) and their recognition relies on the correct identification of image areas corresponding to the same color. However, the problem of color recognition is highly affected by the environment illumination conditions, as well as the robot's camera settings. In this thesis, we propose a new approach to color recognition, which relies on modeling on-line the signatures of the target colors in the color space under different illuminations using density estimation with Gaussian distributions and dynamically identifying and using the correct models by exploiting a metric on the most common color in the RoboCup environment (green). On the problem of field line recognition, we focus on identifying a variety of field line landmarks (straight lines, center circle, corners, T-lines), which are useful for localization. This is accomplished by searching for white pixels in the camera images, selectively keeping those that can be part of a line, and then identifying each type of line using curve fitting techniques. For each recognized line landmark, we use geometry and projection techniques to estimate its distance and bearing with respect to the robot. Our work contributes an on-line tool for color recognition and a real-time module for field line recognition appropriate for on-board execution on the Aldebaran Nao humanoid robots. The proposed methods perform reliably in most cases, failing only in extreme cases, which are typically infrequent during RoboCup games.