Accelerometer and Camera-Based Strategy for Improved Human Fall Detection.

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Human falls pose significant health risks, necessitating reliable detection and classification systems.
• Existing methods often struggle with accuracy and efficiency in real-time fall detection.
• Accelerometric data and human silhouette analysis offer promising features for fall event recognition.

Purpose of the Study:

• To develop and evaluate an anomaly detection strategy for accurate human fall detection and classification.
• To leverage exponentially weighted moving average (EWMA) for fall event identification and feature extraction.
• To utilize Support Vector Machine (SVM) for distinguishing true falls from fall-like events, optimizing classification efficiency.

Main Methods:

• Utilized exponentially weighted moving average (EWMA) monitoring for anomaly detection in accelerometric data.
• Extracted fall-specific features identified by EWMA for subsequent classification.
• Employed Support Vector Machine (SVM) algorithm for binary classification of detected falls versus non-fall events.
• Validated the proposed method on publicly available fall detection datasets.

Main Results:

• The EWMA-based SVM approach demonstrated high accuracy in both detecting and classifying human falls.
• The method successfully distinguished true falls from similar non-fall events.
• Training time and model complexity were reduced by using a subset of data for classification.

Conclusions:

• The proposed anomaly detection strategy effectively detects and classifies human falls.
• This method shows significant potential for integration into early alert systems for fall prevention and response.
• The EWMA-SVM classifier outperformed other common machine learning models, including neural networks, K-nearest neighbors, and naïve Bayes.