Improving Cross-Country Skating Technique with AI-Based Feedback

Sammanfattning

Recent advances in artificial intelligence and computer vision have opened new opportunities for human activity recognition, particularly in sports performance analysis. This thesis explores the use of video-only input for recognizing and evaluating sub-techniques in skating-style cross-country skiing, a sport that has seen limited application of vision-based methods compared to others. While recent technique assessment in cross-country skiing has relied on wearable sensors, these can be complex and intrusive. In contrast, this work investigates whether modern pose estimation models and machine learning techniques can offer an accessible, non-invasive alternative for sub-technique classification and feedback. To investigate this, we developed a multi-stage feedback system based solely on video input. Our first step involved fine-tuning a pose estimation model to better capture skiing-specific movements. The extracted keypoints were segmented into individual cycles, and a dedicated classifier was trained to recognize sub-techniques. These cycles were then matched to expert references using dynamic time warping and rule-based feedback was generated by comparing joint movements and positions between the user and expert. A novel dataset was collected using drone footage, featuring front- and side-view recordings of skiers ranging from beginners to national-level athletes. The fine-tuned pose estimation model achieved 92% accuracy on a test set with two unseen skiers. For gear classification, we reached 93% accuracy on inter-user data using our mixed-level dataset. Feedback generation focused on two common mistakes regarding leg distance and foot flexibility, for which we observed mostly expected behavior and identified current limitations. These results demonstrate the potential of automated feedback systems based on pose estimation. Key challenges remain, including keypoint accuracy, camera variability, person tracking, classification robustness, and dynamic time warping alignment. Despite these, our work highlights the promise of vision-based systems for supporting technique analysis in cross-country skiing and provides a foundation for future research.