Learning to Assess Squat Technique : Video-Based Pose Analysis with Classical and Deep Learning Models

Sammanfattning

This study investigates the use of machine learning models for classifying squattechnique in the context of fitness and rehabilitation. Although machine learninghas been widely applied to human activity recognition, few studies have directlycompared classical and deep learning models for evaluating squat technique usingvideo-based pose data. The main problem addressed is the challenge of accuratelyevaluating squat form through machine learning models, with the aim of improvingtraining and rehabilitation outcomes. The demand for effective and interpretablesolutions is rising as wearable technology and artificial intelligence become moreand more common in the fitness industry. This study compares the performance oftwo machine learning models, Random Forest and a Bidirectional Long Short-TermMemory (BiLSTM) model, for classifying squat technique from video-based posedata. The BiLSTM model effectively captured temporal patterns in sequential posedata, especially when trained on augmented samples. However, the Random Forest,using biomechanically meaningful features, achieved the highest classification accuracyoverall. This study illustrates the strengths and limitations of deep learning andclassical machine learning approaches for squat form classification. These findingsprovide practical insights into how model choice and feature representation impactperformance and interpretability in video-based exercise assessment.