Humanoid Soccer Robot Target Detection and Localization

摘要

A monocular visual target localization method based on You Only Look Once v8 (YOLOv8) and Perspective-3-Point (P3P) for the RoboCup humanoid soccer robot competition is investigated. In this article, we propose a monocular vision-based object localization method for robot soccer competition using "You Only Look Once" v8 (YOLOv8) and Perspective-3-Point (P3P) algorithms. Firstly, the YOLOv8 algorithm is employed to detect the target object within the input image, providing precise positional information. Subsequently, the P3P algorithm is utilized for accurate localization of the detected targets. Considering the inherent instability in the humanoid robot's motion, the Kalman filtering algorithm is further integrated to enhance the accuracy and stability of both target detection and localization results. The proposed target detection and localization methodology was thoroughly evaluated through extensive experimentation on a custom dataset. The obtained results demonstrate a remarkable mean Average Precision (mAP50) of 0.989 for target detection and 0.916 for mAP50-95. Notably, when the robot's camera was positioned at distances ranging from 100 cm to 300 cm from the target object, the maximum average positioning error was measured at 32.87 cm, accompanied by a maximum average relative error of 10.97%. Conversely, at closer distances (< 100 cm), the maximum average positioning error reduced to 2.95 cm, accompanied by a maximum average relative error of 3.69%. Furthermore, the utilization of Kalman filtering yielded a substantial decrease in the maximum difference in distance calculation results between consecutive frames, reducing it from 79.71 cm to 30.02 cm. The experimental findings substantiate the efficacy and resilience of the proposed approach, thereby validating its suitability for addressing the stringent demands of humanoid soccer robots within the competitive domain of RoboCup.