Publication

Abstract : This paper presents a novel and efficient fall detection system designed for real-time monitoring of elderly individuals using an Intel RealSense D455 depth camera integrated with MediaPipe Pose estimation. The proposed method leverages a bounding box-based technique that analyzes feet positions and computes the center of gravity (CoG) from skeletal data to assess posture stability. By relying exclusively on depth sensor-derived skeletal information, the system ensures user privacy by avoiding intrusive video analysis. Quantitative evaluation on recorded video frames yielded a detection accuracy of 90-92% and a latency ranging from 33 to 300 milliseconds, thereby ensuring prompt alerts during potential fall events. Extensive testing demonstrated the system's ability to accurately classify various postures, including sitting, standing, and falls in multiple directions, with the CoG displacement exceeding a critical 50-pixel threshold for fall detection. A comparative analysis with existing wearable sensorbased, machine learning-based, and action recognition approaches revealed that the proposed method offers a compelling balance between efficiency, computational cost, and privacy protection. Moreover, the system maintained robust performance under diverse lighting conditions, affirming its applicability in real-world scenarios such as nursing homes, assisted living facilities, and home care environments. Future work will focus on integrating wearable sensors for cross-verification, implementing adaptive thresholding based on real-time data, and extending the system to predict instability for proactive intervention. These enhancements are expected to further improve detection reliability and broaden the scope of application for vulnerable populations, ultimately contributing to enhanced elderly safety and improved quality of life. Overall, it enhances safety.