Fall Detection Using Commodity Smart Watch and Smart Phone

Ilias Maglogiannis, Charalampos Ioannou, George Spyroglou, Panayiotis TsanakasLecture Notes in Computer Science 2014 · DOI · OpenAlex

Abstract

As the global population ages, the risk of falls among seniors has become a significant public health concern. This paper introduces a real-time, unobtrusive fall detection system that leverages off-the-shelf hardware: a Pebble smartwatch and an Android smartphone. By offloading complex motion data processing to the smartphone and utilizing the Cumulative Sum (CUSUM) algorithm, the system provides a low-cost, energy-efficient solution capable of monitoring users for over 35 hours on a single charge.

Introduction

Approximately one-third of seniors experience a fall annually, often resulting in severe injuries and a loss of independence. While various fall detection systems exist, many are either too intrusive (camera-based) or rely on expensive, specialized medical sensors. The proposed system addresses these challenges by using commodity hardware that users are likely to wear daily, ensuring real-time alerting and high accessibility.

System Architecture

The system employs a dual-module software architecture distributed between the wearable device and a mobile processor.

The Wearable (Pebble Smartwatch)

The smartwatch serves as the primary data source, utilizing its internal 3-axis accelerometer. To optimize battery life and minimize Bluetooth Low Energy (BLE) bandwidth usage, the watch samples motion at 50ms intervals but bundles ten samples into a single string object every 500ms for transmission.

The Processor (Android Smartphone)

The smartphone acts as the "brain" of the system, handling the computational load that early smartwatches could not support. It receives the bundled accelerometer data via the App Message Service and applies the CUSUM change detection algorithm to identify potential falls.

Detailed data flow and processing architecture between the Pebble smartwatch and Android application.

Workflow and Alerting

When the CUSUM algorithm detects a significant change in motion indicative of a fall, the phone triggers a "fall-alarm" on the watch. The watch vibrates and prompts the user to cancel the alarm within five seconds if it is a false positive. If the user does not intervene, the phone automatically executes predefined actions, such as sending an SMS or email to emergency contacts and updating a cloud-based monitoring service.

Experimental Evaluation

The authors conducted initial experiments to calibrate the CUSUM algorithm's sensitivity, governed by the threshold parameter $h$. A lower $h$ value increases sensitivity (faster detection) but leads to more false alarms from activities like sitting down quickly. Conversely, a higher $h$ improves robustness at the cost of a slight detection delay.

Performance metrics showing the trade-off between false alarms and detection delay across different threshold (h) values and axes.

Key findings from the evaluation include:

  • Battery Life: The system achieved over 35 hours of continuous monitoring using adaptive sampling logic.
  • Accuracy: An $h$ value between 3.8 and 5.0 on the X-axis resulted in zero false alarms with a minimal detection delay of 10ms.
  • Axis Sensitivity: The Z-axis was identified as the most susceptible to noise, requiring higher thresholds than the X or Y axes to maintain accuracy.

Discussion and Future Work

While the current system demonstrates the viability of commodity-based fall detection, it faces limitations such as the Pebble SDK's lack of multitasking, which requires the app to remain in the foreground. Future work will focus on integrating contextual data (GPS, audio) to further reduce false positives and estimating fall severity to assist emergency responders.

Summary updated

BibTeX
@inproceedings{maglogiannis2014fall,
  author = {Ilias Maglogiannis and Charalampos Ioannou and George Spyroglou and Panayiotis Tsanakas},
  title = {{Fall Detection Using Commodity Smart Watch and Smart Phone}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2014},
  doi = {10.1007/978-3-662-44654-6\_7},
}
Figure