Depth-based human fall detection via shape features and improved extreme learning machine.

Area of Science:

• Computer Vision
• Biomedical Engineering
• Gerontology

Background:

• Falls are a significant cause of injury in the elderly.
• Wearable fall detection devices are costly and inconvenient.
• There is a need for affordable and user-friendly fall detection solutions.

Purpose of the Study:

• To develop an automated fall detection approach using a low-cost depth camera.
• To combine computer vision techniques for accurate fall identification.
• To optimize the performance of the learning-based classifier.

Main Methods:

• Utilized a low-cost depth camera (Kinect) for data acquisition.
• Extracted Curvature Scale Space (CSS) features from human silhouettes.
• Represented actions as a bag of CSS words (BoCSS) and classified using Extreme Learning Machine (ELM).
• Optimized ELM hyperparameters with a variable-length particle swarm optimization algorithm.

Main Results:

• Achieved up to 91.15% sensitivity, 77.14% specificity, and 86.83% accuracy on a custom dataset.
• Demonstrated comparable performance to multi-camera state-of-the-art methods on a public dataset.
• Successfully distinguished between six daily actions including falling.

Conclusions:

• The proposed depth camera-based system offers an effective and low-cost solution for automated fall detection in the elderly.
• The combination of CSS features, BoCSS representation, and optimized ELM provides robust action recognition.
• This approach presents a viable alternative to traditional wearable sensors for fall monitoring.