YOLOv5-OLCAM-Based Target Detection and RRT-PRM Path Planning for Soccer Robots under Uncertain Conditions

Abstract

Robots still face enormous challenges in soccer matches, as the environment is complex and everchanging. Robots need to perceive the positions and trajectories of teammates, opponents, and the ball in real time. Therefore, based on the You Only Look Once v5 model, an improved object recognition method is designed using a lightweight convolutional attention module. A path planning method is constructed by combining the rapidly-exploring random tree algorithm with the Probabilistic Roadmap method. Finally, a soccer robot control strategy incorporating the rapidly-exploring random tree algorithm is proposed. The research used a ball detection dataset J, specifically designed for the Robot Soccer Standard Platform League for testing. The research results showed that the accuracy and running time of the improved target recognition algorithm under size images were 99.12% and 0.19ms, respectively. The path planning algorithm, integrating the rapidly-exploring random tree algorithm, also performed well, requiring only 800 iterations to obtain the shortest planned path, which was 19.637cm. Compared with other mainstream methods, the improved method had significant advantages in path length and iteration times (p<0.001), indicating its practicality and robustness under uncertain conditions. In the comparison of control strategies, the research method had the lowest global decision entropy of 0.934 and the shortest average planning time of 26.8 seconds. The research method can significantly improve the intelligence level of soccer robots in competitions and assist soccer robots in making optimal control decisions on the field, achieving more efficient collaboration.