Efficient Limb Range of Motion Analysis from a Monocular Camera for Edge Devices.

Area of Science:

• Biomechanics
• Computer Vision
• Medical Technology

Background:

• Traditional limb range of motion (ROM) assessment relies on manual goniometry, which is labor-intensive.
• Advancements in computer vision and deep learning offer potential for automated kinematic analysis.
• Existing deep learning approaches often prioritize accuracy over efficiency and cost, limiting practical clinical application.

Purpose of the Study:

• To develop a high-performance, low-cost, camera-based tool for upper and lower limb ROM measurement.
• To create a lightweight and fast deep learning model optimized for resource-constrained edge devices.
• To balance accuracy with the benefits of edge computing, such as cost-effectiveness and localized data processing.

Main Methods:

• Proposed a lightweight, fast deep learning model for human pose estimation and subsequent limb ROM calculation.
• Employed a compact neural network architecture with 8-bit quantized parameters for improved memory efficiency and reduced latency.
• Evaluated the model on upper and lower limb tasks and tested its performance on a Raspberry Pi edge device.

Main Results:

• The proposed model achieved a 4.1x speed increase and was 15.5x smaller than a state-of-the-art model.
• Demonstrated satisfactory accuracy and agreement with goniometer measurements for limb ROM.
• Experimental results on a Raspberry Pi confirmed maintained accuracy while significantly reducing equipment and energy costs.

Conclusions:

• The developed deep learning model provides an efficient and cost-effective solution for limb ROM assessment.
• Optimization for edge devices enables localized data processing and reduces hardware expenses.
• The approach shows potential for widespread adoption in various healthcare settings and adaptability to diverse hardware environments.