Enhanced Human Fall Detection via Lightweight MDS-OpenPose Framework

Falls among the elderly pose a significant risk of injury and even mortality, underscoring the importance of real-time monitoring systems to mitigate these hazards. Existing posture estimation-based fall detection methods often struggle with high parameter counts, computational complexity, and slow processing speeds. This paper proposes an improved OpenPose algorithm, termed MDS-OpenPose, which addresses these issues. By integrating the lightweight MobileNetV3 network to replace the original VGG feature extraction network, optimizing convolutional layer sizes, and introducing DenseNet dense connections, MDS-OpenPose significantly reduces model complexity while maintaining high accuracy. Fall detection is achieved through a comprehensive method that analyzes vertical distances between the head and feet, trunk tilt angles, and horizontal displacement of the center of mass. Experimental results demonstrate that MDS-OpenPose achieves a substantial improvement in FPS on the COCO dataset while maintaining high precision and recall rates. On the Fall Down dataset, it attains an accuracy of 93.0% and a precision of 92.1%, showcasing its effectiveness and robustness in real-time fall detection applications.