An evaluation of 3D head pose estimation using the Microsoft Kinect v2.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Human-Computer Interaction

Background:

• The Kinect v2 sensor offers a low-cost, accessible solution for real-time 3D head pose estimation.
• Accurate head posture measurement is crucial for assessing postural control, particularly in physical therapy settings.

Purpose of the Study:

• To evaluate the accuracy of the Kinect v2 sensor for non-invasive 3D head pose estimation.
• To determine the sensor's potential as a tool for measuring head posture in clinical and research applications.

Main Methods:

• Comparison of Kinect v2 head pose estimates against a marker-based ground truth system.
• Testing under various head and neck movement conditions with a static torso.
• Analysis of estimation accuracy across different head rotations and distances from the sensor.

Main Results:

• Average errors in absolute yaw, pitch, and roll were 2.0±1.2°, 7.3±3.2°, and 2.6±0.7°, respectively.
• Rotational errors relative to rest pose were 1.4±0.5° (yaw), 2.1±0.4° (pitch), and 2.0±0.8° (roll).
• Estimation failures occurred in 10.2±6.1% of poses during extreme head rotations, but performance was unaffected by distance or lack of face calibration.

Conclusions:

• The Kinect v2 sensor demonstrates significant potential for non-invasive 3D head pose estimation in applications like postural control assessment.
• Its accuracy is sufficient for use in physical therapy, especially when facial features are visible.
• Future developments may involve multi-Kinect setups for comprehensive head and body movement tracking.