DeepFall: Non-Invasive Fall Detection with Deep Spatio-Temporal Convolutional Autoencoders.

Area of Science:

• Computer Science
• Biomedical Engineering
• Artificial Intelligence

Background:

• Human fall detection is critical for health and safety, but the rarity of falls complicates supervised learning.
• Extracting domain-specific features for fall identification is challenging in imbalanced datasets.

Purpose of the Study:

• To present DeepFall, a novel framework for fall detection.
• To formulate fall detection as an anomaly detection problem.
• To leverage deep spatio-temporal convolutional autoencoders for feature learning.

Main Methods:

• The DeepFall framework employs deep spatio-temporal convolutional autoencoders to learn features from normal activities.
• It utilizes non-invasive sensing modalities like thermal and depth cameras.
• A novel anomaly scoring method combines frame reconstruction scores over a temporal window.

Main Results:

• DeepFall was tested on three public datasets using non-invasive sensors.
• The framework demonstrated superior performance in detecting unseen falls compared to traditional autoencoder methods.

Conclusions:

• DeepFall effectively addresses the challenge of fall detection in imbalanced datasets.
• The proposed anomaly detection approach and deep learning model show promise for real-world applications.