Article Preview
TopIntroduction
Winning against the world soccer championship in 2050 is the primary goal in RoboCup. Humanoid robot soccer has to perceive its environment in real-time, mostly using visual information (Gerndt, et al., 2015). New rules update in RoboCup have been made to encourage the robot to adapt to nearly soccer conditions. For example, the soccer ball follows the FIFA standard, at least 50% white. Goalpost color also changes to entirely white, which is above a synthetic grass field. Field lines, goalposts, and the ball nearly have white color, which can deteriorate a conventional detection system such as color segmentation. The real condition of the humanoid soccer league is shown in Figure 1. Because of this change, quite popular methods in the past, such as color segmentation, morphology, and blob detection approaches (Farazi, et al., 2014; Laue, et al., 2009), need further significant improvement.
Figure 1. RoboCup competition in real condition. The green color of synthetic grass on the field is dominant. Ball, goalposts, and lines are in white color. Robot mostly using a dark color as suggested by the rule.
Color segmentation and morphological process together with blob detection are still heavily been used in the RoboCup community. These methods are due to simplicity and real-time running for a most embedded platform. The work by (Härtl, et al., 2013), a color classification performed based on color similarity. The work (Schwarz, et al., 2015) introduced a heuristics approach for developing a calibration-free vision system. The work by (Houliston, et al., 2015) introduced compensation for the change in image illumination by using adaptive search tables. The color intensities in the Y channel of YUV color space are the main focus for (Cano, et al., 2015) to overcome the color changes in the goalpost. Different heuristics also applied to color histograms due to the variations in illumination. In order to increase accuracy, the latest supervised learning algorithm has been performed by (Albani, et al., 2016). However, it requires a quite sophisticated embedded platform to meet real-time detection.
Recent works on humanoid soccer have been performed using deep learning to boost the accuracy of detecting ball (Cruz, et al., 2017; Militão, et al., 2017; Speck, et al., 2016). Using the latest technology of deep learning can improve the ball detector significantly, successfully adapting to the new rule of the ball's pattern. However, a high computation becomes the main problem of employing such an algorithm. It is noted that a humanoid soccer robot has a limited computation power and visual perception task in a robot soccer competition. It is not only detecting ball but also field, lines, goalposts, and its position. Thus, employing all of those visual tasks would not practical.