Instance Segmentation-Based Markerless Tracking of Fencing Sword Tips
Takehiro Sawahata, Alessandro Moro, Sarthak Pathak, Kazunori Umeda · 2024
This study addresses the challenge of detecting the tip of a fencing sword. The swift motion and diminutive size of the fencing sword tip not only poses difficulties in detection but also occasionally lead to its omission from video recordings. Moreover, conventional detection approaches such as affixing markers to the sword tip are unsuitable in sports contexts as they could encumber the athletes. In light of these considerations, our research has devised a system that exclusively employs monocular camera images to consistently gather information about the sword tip. Even in cases where the tip is not captured, we propose a method for predicting its position based on historical data and subsequent interpolation. Specifically, the entire sword is recognized using instance segmentation. And the tip of the sword is identified with skeletal point information. In instances where the tip eludes detection, its position is projected using preceding information and skeletal wrist point data, to ensure uninterrupted tracking. Our proposed method's efficacy was confirmed through various experiments conducted under conditions mirroring actual match scenarios. These experiments demonstrate the effectiveness of our approach.