Home-Based Gait Abnormality Detection Using SAIL-TUG and Machine Learning Techniques

This paper presents a video-based machine learning approach for detecting gait anomalies, focusing on key phases of functional mobility assessments. Designed for home use, the system offers an accessible solution for individuals with Parkinson's disease and other mobility impairments. Our method leverages Random Forest classifiers for gait pattern detection, MediaPipe for pose estimation, and Savitzky-Golay filtering for signal smoothing. Key gait features-such as step cadence, stance times, and hip flexion angles-are extracted to differentiate normal and abnormal gaits. The dataset consists of 404 samples (241 normal, 163 abnormal), with minimal class imbalance. The Random Forest model achieved 88.61 % accuracy and an AUC of 0.89, demonstrating strong performance. Feature analysis highlights cadence, left stance time, and steps before turning as key discriminative factors. Future work will focus on enhancing feature engineering and addressing class imbalance to further improve performance.