In-Bed Pose Estimation: Deep Learning With Shallow Dataset.

Area of Science:

• Computer Vision
• Medical Imaging
• Human Pose Estimation

Background:

• In-bed pose estimation is crucial for healthcare but overlooked in computer vision.
• Existing pose estimation models fail due to unique in-bed challenges like variable lighting and unconventional viewpoints.

Purpose of the Study:

• To develop a robust human pose and body part detection method specifically for in-bed scenarios.
• To address the limitations of current pose estimation models in specialized environments.

Main Methods:

• Proposed an infrared selective (IRS) image acquisition technique for consistent data quality under varying light.
• Introduced a 2-end histogram of oriented gradients (HOG) rectification method to handle unconventional pose perspectives.
• Fine-tuned a pre-trained Convolutional Pose Machine (CPM) model on a custom in-bed IRS dataset.

Main Results:

• The fine-tuned CPM model with HOG rectification achieved a 26.4% improvement in PCK at PCK0.1 compared to the unrectified model.
• The fine-tuned model demonstrated a further 16.6% increase in pose estimation accuracy over traditionally tuned CNNs, even with well-aligned images.

Conclusions:

• The proposed IRS imaging and HOG rectification methods effectively address in-bed pose estimation challenges.
• Fine-tuning pre-trained deep learning models is a viable strategy for specialized pose estimation tasks with limited data.