Deep Learning-Based Approach for Object Detection in Robot Football Competition

Abstract

Robot football competition is a complex and emerging field of artificial intelligence research involving object detection technology, robotics, intelligent control, and other technologies. However, object detection is one of the most core technologies, supporting robots to realize tactical cooperation and real-time actions such as shooting, passing, and obstacle avoidance behavior. With precise accuracy and detection speed requirements, fast-moving robots and footballs must be recognized accurately by object detection algorithms under environments of changing backgrounds and lighting conditions. In this paper, to propose a reliable detection method for robot football competition, an end-to-end training approach is applied based on the YOLOv3 algorithm. K-means reclustering is used to calculate more appropriate bounding box priors to adapt to the size of detected objects. Besides, the smooth L1 loss function is adopted for the loss of the bounding box instead of MSE loss to reduce the model's sensitivity to outliers. With the framework of Pytorch, the proposed method can reach the mAP up to 96.5%, recognizing specific targets under the Standard Platform League(SPL) of Robocup. Accurate object detection algorithms can improve the capabilities of robot behavioral decision-making and positioning. In the future, superior lightweight algorithms can also be deployed on edge devices to meet the visual needs of real-time intelligent services.