Ball Detection and Tracking with Different Embedded Systems in the RoboCup Soccer context

摘要

This paper presents a vision system for ball detection and tracking in the context of the RoboCup Soccer Humanoid League Kidsize category. The system aims to enable humanoid robots to autonomously play soccer by visualizing and interacting with the environment in real time. Previous approaches based on color and shape extraction have limitations due to recent rule changes mandating a partially white ball. To overcome these challenges, we propose the use of convolutional neural networks, specifically the You Only Look Once (YOLO) model, known for its efficiency and real-time performance in object detection. Optical flow, a technique analyzing pixel displacement between frames, is integrated to capture ball movement. Experimental evaluations compare different hardware platforms in terms of execution velocity and computational cost. Our results demonstrate the effectiveness of the developed vision system for ball detection and tracking in the Kidsize category. The YOLO model trained on the TORSO-21 dataset allows for fine-tuning, offering adaptability to changes in context without the need for a complete model overhaul. By incorporating optical flow within the ball area, accurate movement tracking is achieved. The analysis of different hardware platforms provides insights into execution speed, accuracy, computational cost, and the limitations imposed by size constraints. This research contributes to the advancement of computer vision applications in robotics and provides a foundation for further exploration and optimization in the field.