Online Deep Squat Evaluation: Leveraging Subject-Specific Adaptation and Information Retention

Evaluating deep squats accurately during automatic physical rehabilitation monitoring across different subjects remains challenging due to inter-subject variability and limited labelled data. The challenges include: 1) conventional methods presuppose that a "one-model-fits-all"approach works for activity evaluation, ignoring that subject-specific differences can lead to suboptimal results if these differences are not considered. 2) Previous studies focus on offline learning, where models are trained on the entire dataset, which can be updated later through retraining. This approach neglects the need for continual learning, where models adapt sequentially to new subjects while retaining past knowledge to prevent catastrophic forgetting. This study addresses these challenges by proposing a novel continual meta-learning approach and a memory buffer to provide personalized deep squat evaluations. Using Azure Kinect sensors, we collected RGB-D videos and 3D skeletal data from 33 participants performing deep squats, annotated with Functional Movement Screen (FMS) scores. Our model dynamically adapts to new participants while retaining knowledge from previous ones, preventing performance degradation over time. Experimental results demonstrate that our approach outperforms a model without a buffer memory technique by retaining learned knowledge across participants and adapting to new individuals with minimal data.