Fall detection using gait analysis

Chan, Adeline Xyn (2025) Fall detection using gait analysis. Final Year Project, UTAR.

PDF Download (2861Kb)

Abstract

This project focuses on fall detection for elderly populations using deep learning and gait analysis. Falls are a major concern in aging populations, often resulting in severe injuries and diminished quality of life. Traditional fall detection systems have limitations in accurately identifying falls and adapting to real-world environments. This study implements and compares five deep learning architectures: CNN, BiLSTM, BiGRU, CNN-BiLSTM and CNN-BiGRU with attention mechanisms to capture both spatial and temporal patterns in human gait. MediaPipe extracts pose landmarks from video frames, while OpenCV aids in frame processing. The research process involves data collection from the Multiple Cameras Fall Dataset with comprehensive preprocessing including two-stage normalization and time-series scaling. Eight gait features are extracted: stride length, knee angles, body velocity, acceleration, step frequency, posture angle, and arm swing. The data is split using an 80/10/10 sequence-level approach to ensure models are tested on completely unseen video sequences. Performance evaluation using accuracy, precision, recall, F1-score and AUC-ROC metrics shows the hybrid CNN-BiLSTM model achieving 90.04% test accuracy with 94.51% precision and 94.92% recall on completely unseen data. Permutation-based feature importance analysis reveals that arm swing is the most critical predictor across all models, followed by stride length and knee angles. The MediaPipe approach demonstrates a 3.8-fold improvement in processing time compared to traditional raw frame processing while maintaining detection accuracy. The novelty lies in the systematic comparison of five architectures using sequence-level data splitting to prevent data leakage, comprehensive feature importance analysis across multiple models, and the fusion of real-time gait analysis with deep learning techniques. Results demonstrate the hybrid model's ability to detect falls with high accuracy and minimal false detections, providing an efficient and adaptable solution for fall detection in elderly care settings.

Actions (login required)