Algorithms for a fast and reliable color-based global vision for FIRA MiRoSot large league soccer robot system

• Main Author: Gunay, Noel S.
• Format: text
• Language: English
• Published: Animo Repository 2010
• Subjects: Algorithms; Robotics; Artificial Intelligence and Robotics
• Online Access: https://animorepository.dlsu.edu.ph/etd_doctoral/956; https://animorepository.dlsu.edu.ph/context/etd_doctoral/article/1955/viewcontent/CDTG004941_P.pdf
• Institution: De La Salle University

A development of a two-camera soccer robot system (SRS) for the FIRAMiroSot large league is presented in this research. Developed are algorithms that led to a fast and reliable performance of the SRS vision subsystem in recognizing the twenty-three (23) objects of interest using two cameras. This work contributes to the field of color-based multiple object recognition by providing a vision system that reliably identifies and localizes 23 highly dynamic objects of interest within 15 milliseconds even when observed under testing situations that involve collision, rotation, fast motion, shadow influence and prolonged system operation all under inconsistent illumination. A new color uniform is designed and a fast algorithm used to identify the eleven home robots wearing the uniform is developed and tested of its performance through real experiments. Also developed is a library that is primarily intended for a fast high level color-based object recognition that makes use of the OpenCV data structures and highly optimized functions. Another algorithm developed which is also aimed to reduce the overall processing time is an incremental tracking approach that makes a selective use of a linear prediction method which is complemented by a rule-based predictor of collision to boundaries and backed-up by the use of a second search window and then of the postponed shifting to next frame of the full-tracking operation as the ultimate resort in recovering a “lost” object. Using this algorithm adds the velocities and predicted positions of the 23 objects to the usual vision data produced by existing vision systems. For two-camera operation, an algorithm that eliminates the redundant full tracking of the overlap region when the vision system operates in incremental tracking mode is also developed purposely to also reduce the overall processing time. The effectiveness of this algorithm is tested using a devised measure for using incremental tracking which could serve as a new benchmark. A webcam and firewire implementations of the SRS are developed and were used to test the developed algorithms. The overall processing time obtained in recognizing the 23 objects is 21.68 milliseconds when the vision system operates in full tracking mode and about 15 milliseconds when in incremental tracking mode at an almost 100% effectiveness which means that the vision system is in incremental tracking mode most of the time, and this effectiveness is only reduced by the number of times the ball gets occluded or in very rare occasions when the number of opponent robots detected falls below some threshold value. This overall processing time of 15 milliseconds improved on the 22 milliseconds of an existing work which also uses the MiroSot large league as the test bed. All the 23 objects were reliably identified in all experiments under laboratory lighting conditions that span a wider range of differing illumination similar to what can be expected in different competitions. ii