Continuous detection of human fall using multimodal features from Kinect sensors in scalable environment

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automatic human fall detection is crucial for video surveillance and home monitoring systems.
• Existing unimodal methods (RGB, depth, skeleton) face limitations due to lighting conditions and reliability issues.
• Previous fall detection systems were often restricted to small spaces and offline video streams.

Purpose of the Study:

• To develop a robust human fall detection system overcoming limitations of unimodal approaches.
• To combine multi-modal features (skeleton and RGB) for enhanced detection accuracy and reliability.
• To enable real-time monitoring in large environments using a scalable architecture.

Main Methods:

• A hybrid approach combining skeleton-based rules (vertical velocity, height) and RGB-based motion analysis (motion map, kernel descriptor, Support Vector Machine).
• Utilized Kinect sensor data, integrating skeleton and RGB features to leverage the strengths of each modality.
• Implemented a client-server architecture with late fusion techniques for deploying the system across multiple Kinects in large areas.

Main Results:

• Achieved higher accuracy and a lower false alarm rate compared to existing methods on public datasets.
• Demonstrated successful online validation in a large lab environment using multiple Kinects.
• Attained an average accuracy of 91.5% at a frame rate of 10 frames per second.

Conclusions:

• Multi-modal feature integration significantly outperforms unimodal approaches for human fall detection.
• The developed system's online deployment capability highlights its potential for real-world applications in various living spaces.
• The method offers a scalable and reliable solution for continuous monitoring and fall incident detection.