Spatial–Temporal Transformers With Stochastic Time-Warping and Joint-Wise Encoding for Rehabilitation Exercise Assessment

บทความในวารสาร

ผู้เขียน/บรรณาธิการ

กลุ่มสาขาการวิจัยเชิงกลยุทธ์

รายละเอียดสำหรับงานพิมพ์

รายชื่อผู้แต่ง: Matangkasombut, T.; Kumwilaisak, W.; Hansakunbuntheung, C.; Thatphithakkul, N.

ปีที่เผยแพร่ (ค.ศ.): 2026

URL: https://www.scopus.com/inward/record.uri?eid=2-s2.0-105026472061&doi=10.1109%2FOJCS.2025.%603650355&partnerID=40&md5=8011c945083aba486ae02a6698a3e52b

ภาษา: English-Great Britain (EN-GB)

ดูบนเว็บไซต์ของสำนักพิมพ์

บทคัดย่อ

Accurate and objective assessment of rehabilitation exercises is critical for ensuring correct execution and maximizing patient recovery, particularly in unsupervised or home-based settings. Existing deep learning approaches frequently rely on graph-based skeletal representations with predefined topologies, which constrain the discovery of long-range or task-specific joint dependencies and limit adaptability across datasets with varying skeletal definitions. To address these limitations, we propose a Spatial–Temporal Transformer framework that directly models 3D joint position data without requiring an explicit adjacency matrix. The framework incorporates a joint-wise feature encoding and structure embedding mechanism to provide unique representations for each joint, thereby mitigating ambiguities arising from symmetry or overlapping movements. Furthermore, a stochastic time-warping augmentation strategy is introduced to simulate execution speed variations, enhancing robustness to diverse patient movement patterns. By applying small, randomized temporal scaling to local segments while consistently interpolating spatial coordinates within temporal boundaries, this stochastic variation enriches the dataset significantly while preserving the biomechanical patterns. Experimental results on the KIMORE dataset demonstrate that the proposed method reduces mean absolute deviation (MAD) by 67.4 % relative to the current state of the art, while also maintaining strong generalization on the UI-PRMD dataset. The approach is compatible with multiple pose estimation algorithms and acquisition modalities, making it suitable for deployment in real-world telerehabilitation and clinical monitoring applications. © 2020 IEEE.

คำสำคัญ

ไม่พบข้อมูลที่เกี่ยวข้อง