PersonalPT: One-shot approach for skeletal-based repetitive action counting for physical therapy

Abstract

There are thousands of physical therapy exercises which can be selected to tailor an individual’s rehabilitation program. In addition, exercises can be modified to accommodate a patient’s strength and range of motion as they recover and progress. The large size of the resulting set of exercises and their variations is problematic for current evaluation and feedback techniques, which are trained on a small number of exercises. Real-time exercise repetition counting, a core functionality for automated exercise feedback, is useful for promoting better health outcomes for physical therapy patients performing at-home exercises. We propose PersonalPT, a smartphone-based solution which can be used by physical therapists to customize individual patient treatment plans with a single training example. Our proposed one-shot exercise repetition segmentation model allows physical therapists to enable repetition counting on any exercise for individual patients based on their physical ability and rehabilitative needs. Our machine learning model outperforms other repetition counting algorithms (another semi-supervised and a supervised approach) on three exercise datasets. We demonstrate the feasibility of using computer vision and machine learning, on a smartphone, to perform repetition counting for exercises in real-time.