Thermal Image Processing for Respiratory Estimation from Cubical Data with Expandable Depth

Area of Science:

• Biomedical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Rising healthcare costs necessitate affordable, non-invasive vital sign monitoring.
• Respiratory rate (RR) is a crucial health indicator, but current estimation methods are complex.
• Existing RR estimation techniques often require multiple processing steps, limiting deployment on edge devices.

Purpose of the Study:

• To develop a single neural network for end-to-end respiratory rate estimation.
• To enable RR estimation in a single forward pass, suitable for resource-constrained edge devices.
• To overcome limitations of existing multi-step image and signal processing methods.

Main Methods:

• A multi-path neural network architecture leveraging video recognition.
• Capturing both spatial and temporal information from thermal video data.
• Processing data at different sampling rates to analyze temperature variations in the nostril area during breathing.

Main Results:

• The end-to-end solution achieved superior performance compared to state-of-the-art methods.
• Demonstrated robustness on low-resolution thermal video sequences.
• Eliminated the need for additional pre/post-processing and signal-processing techniques.

Conclusions:

• The proposed single neural network offers an efficient and convenient method for RR estimation.
• The approach is suitable for deployment on embedded edge systems with size and power constraints.
• This technology has the potential for widespread application in local, user-close monitoring solutions.