Microsoft Azure Kinect Calibration for Three-Dimensional Dense Point Clouds and Reliable Skeletons

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Multi-camera systems are increasingly vital for applications like localization, mapping, and human activity recognition.
• Precise camera calibration is a prerequisite for high-accuracy multi-camera applications.
• Azure Kinect RGB-D sensors offer advanced capabilities but require robust calibration for optimal performance.

Purpose of the Study:

• To analyze existing two-camera calibration methods.
• To propose a guideline for calibrating multiple Azure Kinect RGB-D sensors.
• To achieve optimal alignment of point clouds and skeletal joints.

Main Methods:

• Exploration of 2D and 3D calibration methodologies using Azure Kinect functionalities.
• Implementation of various calibration approaches for RGB-D sensors.
• Evaluation of calibration accuracy for color and infrared point clouds and skeletal data.

Main Results:

• 3D calibration procedures yield superior point cloud alignment compared to 2D methods.
• Average point cloud distances achieved with 3D calibration: 21.426 mm (color, static), 9.872 mm (infrared, static), 20.868 mm (color, dynamic), 7.429 mm (infrared, dynamic).
• Optimal skeletal joint alignment is achieved using 3D calibration on infrared images, with an average error of 35.410 mm.

Conclusions:

• 3D calibration is the recommended approach for multi-Azure Kinect RGB-D sensor systems.
• Infrared sensor data combined with 3D calibration provides the best skeletal joint alignment.
• The proposed guidelines enhance the precision of multi-camera systems for various applications.