本研究旨在解決居家與社區復健場景中缺乏治療師即時監督所導致之代償動作偵測問題。患者在執行復健動作時，常因肌力不足、關節活動度受限或疼痛而透過軀幹偏移、聳肩、肘彎曲或左右受力不均等方式完成動作，若僅判斷動作是否完成，容易忽略錯誤動作模式並增加二次傷害風險。為此本研究提出一套結合幾何規則與視覺語意模型之復健代償動作偵測系統，以 YOLOv11 Pose 或骨架座標擷取人體關鍵點，計算肩關節角度、肘關節角度、軀幹傾斜、肩線旋轉、手部軌跡與左右對稱性等幾何特徵，作為代償判斷之客觀依據；同時導入 SmolVLM 作為受幾何事實約束之 Template 分類器，根據 rest、start、mid 與 max 多張關鍵影像進行多影像語意對齊，輸出對應 template_id，再由系統查表產生穩定之原因說明與改善建議。此外本研究加入 RAG 相似案例檢索、fallback 一致性檢查與個人化 ROM 校準，以降低小樣本資料、模型幻覺與患者個體差異造成的誤判。實驗結果顯示，在自建影像資料集 180 筆樣本中，完整 Geometry + SmolVLM 方法達到 90.56% Accuracy、93.22% Precision、90.83% Recall 與 92.01% F1-score，優於 Geometry only 之 85.06% F1-score 與 SmolVLM only 之 55.71% F1-score；消融實驗亦證明 fallback 與個人化校準能提升系統穩定性與泛化能力。研究成果顯示，本系統可在保留幾何可解釋性的同時提供語意化復健回饋，具備應用於居家復健、臨床輔助評估與智慧復健教練系統之潛力。;This study aims to address the problem of compensatory movement detection caused by the lack of real-time therapist supervision in home-based and community rehabilitation settings. When performing rehabilitation exercises， patients may complete the required movement through trunk deviation， shoulder elevation， elbow flexion， or asymmetric weight bearing due to muscle weakness， limited range of motion， or pain. If the system only determines whether the movement is completed， incorrect movement patterns may be overlooked， thereby increasing the risk of secondary injury. To address this issue， this study proposes a rehabilitation compensatory movement detection system that integrates geometric rules and a vision-language model. YOLOv11 Pose or skeletal coordinates are used to extract human keypoints， and geometric features such as shoulder joint angle， elbow joint angle， trunk inclination， shoulder-line rotation， hand trajectory， and bilateral symmetry are calculated as objective evidence for compensatory movement detection. In addition， SmolVLM is introduced as a geometry-constrained template classifier. Based on multiple key images， including rest， start， mid， and max stages， SmolVLM performs multi-image semantic alignment and outputs the corresponding template_id. The system then retrieves stable explanations and corrective suggestions through a template mapping mechanism. Furthermore， RAG-based similar case retrieval， fallback consistency checking， and personalized ROM calibration are incorporated to reduce misjudgments caused by limited training samples， model hallucination， and individual patient differences. Experimental results show that， on a self-built image dataset containing 180 samples， the complete Geometry + SmolVLM method achieved 90.56% Accuracy， 93.22% Precision， 90.83% Recall， and 92.01% F1-score， outperforming Geometry only with 85.06% F1-score and SmolVLM only with 55.71% F1-score. The ablation study further demonstrates that fallback and personalization improve system stability and generalization. The results indicate that the proposed system can provide semantic rehabilitation feedback while preserving geometric interpretability， showing its potential for home rehabilitation， clinical assistant assessment， and intelligent rehabilitation coaching systems.
