Skill Level Recognition of Recreational Alpine Skiers

The goal of continuous, unobtrusive monitoring and automatic skill recognition of recreational alpine skiers is ambitious; however, research in this area remains in its early stages. This thesis addresses this gap by: (i) designing a smartphone-based sensing pipeline centered on a novel Skier Fixed Reference Frame (SFRF); (ii) developing an unsupervised segmentation and asymmetric multi-channel autoencoder-based multi-task learning (AE-MTL) algorithm for real-world skiing activity and style recognition; and (iii) introducing a consistency score to quantify skier performance and assess skill levels.
The primary validation on data collected from a full-body sensor setup, including two smartphones, demonstrates that the suggested preprocessing algorithm, based on SFRF, invariably detects skiing turns with an overall RMSE of 0.77 and MAE of 0.50 independent of sensor placement and orientation relative to a reference sensor. Further evaluation on a real-world dataset of 18 participants performing six skiing turn techniques shows 96% accuracy in activity detection, robustness to sensor and device placement variation, and 94% leave-one-subject-out accuracy in skill recognition. Crucially, a skill-related distribution shift is identified and quantified: low-skilled and experienced skiers occupy distinct regions in latent feature space. This insight explains why the AE-MTL model, despite achieving state-of-the-art results on standard HAR benchmarks, struggles to generalize across mixed-skill datasets. Moreover, the additional filtering experiment confirms that consistency score variability stems from behavioral noise rather than sensor artifacts.
Together, these contributions establish the first end-to-end, smartphone-only system for recreational skiing motion analysis, laying the groundwork for scalable, data-driven tools that provide real-time feedback and promote lifelong skill development.