Evaluating Feature-Based Homography Pipelines for Dual-Camera Registration in Acupoint Annotation

Area of Science:

• Medical Imaging
• Computer Vision
• Traditional Korean Medicine

Background:

• Acupoint localization is critical for AI/XR in traditional Korean medicine.
• Current 2D image annotation methods exhibit significant inter- and intra-annotator variability.
• Standardized and accurate acupoint identification is needed for technological advancements.

Purpose of the Study:

• To develop a low-cost, dual-camera imaging system for reliable acupoint localization.
• To fuse infrared (IR) and RGB views for enhanced point visibility and standardization.
• To evaluate feature detection and matching algorithms for robust homography estimation on a Raspberry Pi 5 platform.

Main Methods:

• A dual-camera system integrating IR and RGB imaging on a Raspberry Pi 5.
• Utilized an IR ink pen and 780 nm emitter array for standardized acupoint marking.
• Evaluated five feature detectors (SIFT, ORB, KAZE, AKAZE, BRISK) with two matchers (FLANN, BF) for homography pipelines.
• Assessed system performance across varying camera distances and hand postures.

Main Results:

• IR ink demonstrated superior contrast and visibility under IR illumination for acupoint detection.
• KAZE + FLANN achieved the lowest mean 2D error (1.17 ± 0.70 px) and aspect-aware error (0.08 ± 0.05%).
• Dual-camera registration maintained mean 2D error below ~3 px and aspect-aware error below ~0.25% across different postures.

Conclusions:

• The proposed dual-camera system offers a practical and cost-effective solution for high-quality acupoint dataset generation.
• This framework supports the development of AI-based localization and XR integration in traditional Korean medicine.
• The system demonstrates stable, reproducible performance suitable for automated acupuncture education systems.