Noise-robust markerless video gait anomaly detection via two-stage acquisition and LSTM autoencoders

Abstract

We propose a markerless gait anomaly detection system for orthopedic screening. The framework combines MediaPipe-based joint tracking, unsupervised LSTM-autoencoder modeling, and targeted preprocessing to address clinical video noise. Trained only on normal gait, the model detects abnormal patterns in sarcopenia (SA) and Parkinson's disease (PD) patients, achieving detection rates of 97% and 88%, respectively. Our method achieves state-of-the-art performance in sarcopenia detection, surpassing recent sensor-based approaches that require wearable devices or handcrafted features. Furthermore, to the best of our knowledge, this is the first unified markerless framework capable of identifying both sarcopenia and Parkinson's disease using a single video-based system. To improve input quality, we address three common sources of error-frame imbalance, clothing interference, and background clutter-using YOLO-based frame filtering and semantic segmentation. This increased usable gait data by up to 38%. Joint-level analysis identified the knees as the most responsive to gait abnormalities, enabling interpretable and localized assessments. Our results highlight the potential of a scalable, non-invasive system for early detection and monitoring of musculoskeletal disorders.