On-Line Fall Detection via Mobile Accelerometer Data

Stavros V. Georgakopoulos, Sotiris K. Tasoulis, Ilias Maglogiannis, Vassilis P. PlagianakosIFIP 2015 · DOI · OpenAlex

Abstract

Falls represent a significant health risk for the elderly, necessitating reliable and immediate detection systems. This paper proposes a lightweight, real-time framework for detecting falls using 3-axis accelerometer data from standard smartphones. By integrating Candid Covariance-free Incremental Principal Component Analysis (CCIPCA) for dimensionality reduction with the Cumulative Sum (CuSum) algorithm for statistical change detection, the system achieves high accuracy and low latency. The approach is computationally efficient, making it suitable for continuous on-device monitoring without significant impact on battery life or memory resources.

Introduction

Falls are a leading cause of injury and loss of independence for individuals over 65, with approximately one-third of this demographic experiencing a fall annually. While various monitoring solutions exist—including acoustic sensors and computer vision—they are often intrusive, expensive, or limited to specific environments. Smartphones provide a ubiquitous alternative, yet they introduce technical challenges: raw 3-axis accelerometer data is inherently noisy and sensitive to device orientation, and mobile hardware imposes strict resource constraints on processing power and battery consumption.

Raw 3-axis accelerometer data showing the complexity and noise inherent in mobile sensor signals.

Methodology

The proposed framework utilizes a two-stage pipeline designed for streaming data processing:

  1. Feature Extraction via CCIPCA: To address the variability of phone orientation, the system employs Candid Covariance-free IPCA. Unlike standard PCA, which requires batch processing, CCIPCA updates the principal direction of movement incrementally with every new data point (approximately every 12ms). This reduces the 3D signal into a 1D "most significant" signal that is invariant to the phone's rotation.
  2. Change Detection via CuSum: The resulting 1D signal is processed by the Cumulative Sum (CuSum) algorithm. CuSum calculates the log-likelihood ratio between normal walking ($H_0$) and a fall event ($H_1$). An alarm is triggered when the cumulative score exceeds a predefined threshold ($h$).
The transformation from the 1D projected sensor signal (left) to the CUSUM decision function (right), highlighting the detection point.

Experimental Setup

Researchers evaluated the framework using a custom dataset captured from a standard smartphone placed in the pockets or purses of volunteers. Participants performed sequences of normal walking and intentional falling. Data was sampled at 12ms intervals and partitioned into training sets (to establish baseline movement means) and test sets for performance evaluation.

Key Results and Discussion

The study demonstrates that the IPCA-based projection significantly outperforms individual axes or standard batch PCA in several metrics:

  • Accuracy and Reliability: The IPCA method captures the most energetic component of movement regardless of phone orientation. In tests where the Y-axis failed to detect falls, the IPCA projection maintained high reliability.
  • False Alarm Suppression: At a threshold of $h=0.200$, the IPCA method achieved zero false alarms. In contrast, using the X-axis alone resulted in one false alarm, and the Z-axis produced five.
  • Real-Time Performance: The incremental nature of the algorithm allows for minimal detection delay. At lower sensitivity thresholds, the delay between a fall and its detection was as low as a single timestep (~12ms).

One critical insight is the trade-off governed by the threshold parameter $h$. While lower thresholds provide near-instant detection, they increase the risk of false positives. The IPCA projection provides a wider "sweet spot" for this parameter compared to raw sensor axes, enhancing the system's robustness.

Conclusion

This research provides a viable path for non-intrusive, on-line fall detection using existing mobile hardware. Future work will focus on automated parameter tuning and sensor fusion—such as incorporating microphone data to detect impact sounds—to further distinguish between actual falls and benign events like dropping the device. Additionally, the authors intend to validate the framework against larger, public datasets to ensure performance across diverse demographics and hardware specifications.

Summary updated

BibTeX
@inproceedings{georgakopoulos2015line,
  author = {Stavros V. Georgakopoulos and Sotiris K. Tasoulis and Ilias Maglogiannis and Vassilis P. Plagianakos},
  title = {{On-Line Fall Detection via Mobile Accelerometer Data}},
  booktitle = {IFIP},
  year = {2015},
  doi = {10.1007/978-3-319-23868-5\_8},
}
Figure