Pole-arina: Deep Learning–Based Coaching System for Pole Dancing Technique

This thesis presents Pole-Arina, a marker-less coaching system for static pole dancing tricks that analyzes training videos to recognize the performed trick and grade the final pose with transparent, geometry-based feedback. A domain-specific dataset was curated and annotated for this purpose.It includes 836 clips from 58 participants, labeled with a state scheme that supports multi-trick recognition and explicit background modeling.Pole-Arina combines a lightweight bidirectional LSTM for frame-wise recognition with a rule engine that evaluates trick-specific orientations, joint alignments, and proximities, rendering interpretable overlays and concise tips. The model achieved 93.82% per-frame accuracy across all classes and 98.74% trick-only accuracy on end-pose frames.A controlled between-groups user study compared Pole-Arina against traditional video self-review.Participants using Pole-Arina reported significantly higher trust in the feedback and greater clarity for how to improve, and rated usability higher. These results indicate that Pole-Arina can deliver accurate recognition and actionable feedback that users trust and understand, making structured coaching accessible outside the studio. This work establishes a practical baseline for AI coaching in pole sports.