کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Interpretable Skeleton-Based Deep Learning for Automated Squat Quality Assessment in Rehabilitation

نویسندگان :

Mohammad Rezaei¹ Shaghayegh Mohammadikhaveh² Mohammad Zarei³ Hadi Amiri⁴

1- دانشگاه تهران 2- دانشگاه علوم توانبخشی و سلامت اجتماعی 3- دانشگاه تهران 4- دانشگاه تهران

کلمات کلیدی :

Rehabilitation engineering،Human motion analysis،Time series classification،Skeleton-based deep learning،Explainable artificial intelligence

چکیده :

Reliable feedback on exercise form is essential in clinical rehabilitation, but continuous in-person supervision is rarely possible. This paper presents an automatic squat quality assessment approach using 3D skeleton time series from the KIMORE (Ex5) recordings. Each repetition is aligned to a fixed length and converted into a compact motion representation based on bone direction vectors and their temporal derivatives, so the model learns movement quality rather than camera position or body size. We evaluate two deep sequence models: a CNN–BiLSTM hybrid and a Temporal Convolutional Network (TCN) with dilated temporal filters. To improve robustness, we use realistic time-series augmentation and probability averaging with test-time augmentation (TTA). Beyond classification performance, we apply a Grad-CAM-style analysis to produce an importance curve over time, showing which movement phases most influenced the decision, which is valuable for clinical trust and practical feedback. In our experiments, the TCN achieved 80.4% accuracy and 79.9% balanced accuracy on the test set, outperforming the CNN–BiLSTM baseline.