Skeletons distance evaluation for automated control of exercises in physical therapy

Abstract. The paper focuses on developing a method for evaluating the similarity between skeletal models of an instructor and a patient during physical therapy exercises. Unlike general-purpose measures, our approach considers the specific characteristics of therapeutic exercises, including initial positioning, predominant movements, and exercise pace. By incorporating the concept of informativeness for skeletal points and refining normalization techniques, we improve the accuracy of pairwise dissimilarity measures. Experimental results demonstrate improved separability of records based on the accuracy of exercise repetition, highlighting the potential of this approach for enhancing automated physical rehabilitation systems.