Estimating infant upper extremities motion with an RGB-D camera and markerless deep neural network tracking: A...

Area of Science:

• Biomedical Engineering
• Developmental Pediatrics
• Computer Vision

Background:

• Quantitative biomarkers of infant motion are crucial for predicting movement disorders.
• Current methods for motion tracking can be costly or complex.
• Markerless motion analysis offers a potential low-cost alternative.

Purpose of the Study:

• To present and validate a low-cost, markerless motion tracking method for estimating infant upper body kinematics.
• To extract quantitative biomarkers from infant motion for movement disorder prediction.
• To assess the accuracy of the proposed method using a validated setup.

Main Methods:

• Utilized a single RGB-D camera and the DeepLabCut 2D motion tracking software.
• Developed an algorithm to generate 3D coordinates from 2D points, handling missing data.
• Validated the method with a doll on a rotating turntable, testing two camera orientations.

Main Results:

• Evaluated upper body segment lengths, joint angles, and linear point velocity against known values.
• Determined Mean Absolute Errors (MAE) for each kinematic measure, indicating expected margins of error.
• The 45° camera angle simulation highlighted challenges with occlusions, relevant to live infant recordings.

Conclusions:

• The validated markerless motion tracking method provides a feasible approach for obtaining quantitative biomarkers from infant upper body motion.
• The identified MAE offers a benchmark for expected accuracy in clinical applications.
• Future improvements in hardware and tracking software can further reduce error margins for enhanced predictive capabilities in movement disorders.