Abstract
Falls represent a significant health risk for elderly populations, particularly those living alone. This research introduces a multi-modal fall detection system that integrates 3D accelerometer data with overhead video tracking. By fusing these perceptual components and employing Support Vector Machine (SVM) classification, the system achieves detection accuracies between 98.7% and 100%. Furthermore, the system incorporates a severity estimation feature based on post-fall mobility, enabling prioritized emergency responses.
Introduction
For elderly individuals, the primary danger of a fall is often the "long lie"—remaining incapacitated on the floor for extended periods without assistance. Existing automated detection systems typically rely on a single modality, such as wearable sensors or video monitoring. However, sensors can be prone to false alarms from sudden non-fall movements, while video systems often struggle with occlusions or the "backgrounding" of stationary subjects. This paper addresses these limitations through a fused architecture that uses video data to verify sensor-detected incidents, creating a more reliable fail-safe mechanism for home monitoring.
System Architecture and Methodology
The proposed system utilizes two primary data streams to monitor subjects and classify activities:
- Motion Sensing: Subjects wear Sentilla Perk nodes equipped with 3D accelerometers on the chest and belt. These nodes transmit data via the ZigBee protocol to a central monitoring unit.
- Video Tracking: Overhead cameras monitor the environment using a dynamic foreground segmentation algorithm based on Stauffer’s Gaussian Mixture Model (GMM). To prevent the system from losing track of a fallen, immobile subject—a common failure point in computer vision—the authors integrated Kalman filters to adapt learning rates spatiotemporally.
The data from both streams are fed into a Support Vector Machine (SVM) classifier implemented via the WEKA tool. The feature vector includes X, Y, and Z acceleration values alongside bounding box coordinates (position, width, and height) derived from the video tracker.
Experimental Results
The system was validated through controlled experiments involving four distinct fall scenarios (labeled FallA through FallD) and various daily activities like walking. Using 10-cross fold validation, the SVM models demonstrated exceptional performance across all test cases.
Key findings include:
- High Detection Accuracy: Combining (interlacing) data from both the chest and belt sensors resulted in accuracy rates of 98.7% to 100%, consistently outperforming single-sensor configurations.
- Severity Estimation: The system successfully estimated incident severity by calculating the standard deviation of movement for 15 seconds post-fall. High post-fall movement indicates a likely recovery, while near-zero movement suggests potential unconsciousness, allowing the system to prioritize emergency alerts.
Discussion and Future Work
The primary innovation of this work is the "verification" capability provided by data fusion; visual data effectively filters out false positives that might be triggered by an accelerometer spike alone. Additionally, the use of Kalman filters solves the common computer vision problem where stationary subjects are merged into the background model.
Despite these results, the study's limitations include a small test group and the requirement for multiple wearable sensors and overhead cameras, which may present privacy or installation hurdles. Future research should focus on larger-scale testing in diverse home environments and the potential integration of additional modalities, such as audio, to further refine detection capabilities.
Summary updated