Three-Dimensional Sphere Recognition and Tracking Based on YOLO

Abstract

Traditional art exhibitions are usually dominated by relatively static displays such as text, pictures and common multimedia technology. Subject to technical limitations, the exhibition means are relatively simple and the content is relatively thin, which cannot fully meet the exhibition needs of the organizers, nor can it mobilize the enthusiasm of the visitors, and fails to fully show the communication of the exhibition. Therefore, an object detection model based on You Only Look Once(YOLO) network is proposed in this paper to recognize and track the spheres made by felt process. First, the YOLO network was pre-trained using the open source data set, and then the pre-training model was fine-tuned according to the felt sphere image training set. Before fine tuning, the k-means clustering algorithm was used to cluster the marking information of the sphere training set made by felt process. Secondly, for the display of the effect after recognition, OpenCV image processing is used for image special effect processing of the specific recognition area. Through the experimental results, the object detection based on YOLO network proposed in this paper can reach 80.95% in detection accuracy mAP@0.5:0.95 and detection speed up to 20ms, showing excellent performance in detection accuracy and detection speed. It can fit the background interactive display effect of felt art well.