| Title |
Dynamic Frame Sampling-based Sequence Encoding for Video Curling Shot Classification Model |
| Authors |
손유성(Yuseong Son) ; 박재영(Jaeyoung Park) ; 전병환(Byunghwan Jeon) |
| Keywords |
Curling shot classification; Dynamic sampling; Sequence encoding; AI in sports |
| Abstract |
This study proposes a video-based deep learning framework for automatically classifying curling shot types from real match footage. Curling shots are characterized by temporal dynamics from stone motion and collisions to the final resting state. Therefore, shot intent and outcomes are difficult to determine reliably using single-image analysis alone, and manual video analysis is costly and prone to subjective variability. To address these challenges, we developed a shot classification model using video sequences acquired in real-world game environments. The dataset consists of top-view videos extracted from players’ training footage collected at the Uiseong Curling Center in Gyeongbuk and the Pyeongchang Curling Center in Gangwon. A single shot was defined as the interval from the moment the delivered stone crosses the hog line until all stones come to rest. In total, 763 shots were labeled into five classes (Double Take-out, Draw, Guard, Hit & Stay, and Take-out). For the input, sequences were constructed to include informative segments using an importance score derived from difference images that reflect temporal changes across frames. Spatial features were extracted from each frame using a ResNet-based CNN, and temporal dynamics were modeled with a Temporal Transformer. In addition, normalized metadata, namely the throwing order (turn) and whether the team has the hammer, were fused for final classification. Using stratified 3-fold cross-validation and one-vs-rest ROC analysis, the class-wise AUC ranged from 0.773 to 0.938, with the Guard class achieving the highest separability at 0.938±0.014. These results quantitatively demonstrate the feasibility of real match video-based curling shot classification and provide a baseline model and data construction methodology for future coaching systems and automated training feedback applications. |