Human Fall Detection with Ultra-Wideband Radar and Adaptive Weighted Fusion.

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Human fall detection is challenging due to the high similarity between different fall types.
• Existing methods struggle to accurately distinguish between various fall actions.
• Radar imaging offers a promising, non-invasive approach for fall monitoring.

Purpose of the Study:

• To develop an advanced human fall classification system.
• To improve the accuracy of distinguishing between similar fall types using radar data.
• To introduce an innovative deep learning architecture for fall recognition.

Main Methods:

• Designed the SE-Residual Concatenate Network (SE-RCNet) with Squeeze-and-Excitation (SE) modules.
• Utilized SE-RCNet for classifying three types of radar images: time-distance, time-distance, and distance-distance images.
• Implemented an adaptive weighted fusion strategy to combine classification results from different radar image types.

Main Results:

• SE-RCNet achieved high F1-scores for individual radar image types (94.0%, 94.3%, 95.4%) on a self-built dataset.
• Adaptive weighted fusion significantly improved the overall fall classification accuracy.
• The fused system reached a final F1-score of 98.1% for human fall classification.

Conclusions:

• The proposed SE-RCNet with adaptive weighted fusion effectively addresses the challenge of distinguishing similar human fall types.
• Radar image classification using the SE-RCNet architecture demonstrates high performance in fall detection.
• This approach offers a robust and accurate solution for human fall monitoring systems.