Deep Learning-Based Classification of Anterior Pelvic Tilt Rehabilitation Exercises Using RGB Video and Pose Estimation

Abstract

Physical rehabilitation plays an important role in correcting imbalances in the skeletal system, such as anterior pelvic tilt (APT). APT is a medical condition that might become a trigger for developing lower back pain and abnormal posture. Traditional rehabilitation methods involve in-person supervision. The in-person supervision shrinks the scalability and imposes serious limitations on accessibility. This research describes a deep learning-based approach for classifying APT rehabilitation exercises using RGB videos. A custom dataset of over 1000 annotated videos was used to train models. Each video corresponding to one of five core exercises performed with the help of professional physiotherapists (about 150 videos per exercise including variations). Pose estimation was conducted using MediaPipe to extract 3D joint landmarks from each video frame. These pose sequences were fed into multiple deep learning architectures, including LSTM, Transformer, Conformer, Temporal Fusion Transformer (TFT), and CNN-LSTM with attention. The Transformer and TFT models achieved the highest accuracy (96.13%). This work lays the foundation for scalable, camera-based physiotherapy systems capable of delivering real-time feedback for home-based rehabilitation.