A Novel Multiple Camera RGB-D Calibration Approach Using Simulated Annealing

Abstract

The development of a cost-effective surface scanning system tailored for live animal image capture can play an important role in biomedical research. The primary aim was to introduce a low-cost system, achieving a surface reconstruction error of less than 2mm, and enabling rapid acquisition speeds of approximately 1 second for a complete 360-degree surface capture. Leveraging a five RGB-D camera configuration, our approach offers a simple, low-cost alternative to conventional lab-based 3D scanning setups. Key to our methodology is a novel calibration strategy aimed at refining intrinsic and extrinsic camera parameters simultaneously for improved accuracy. We introduce a novel 3D calibration object, extending existing techniques employing ArUco markers, and implement a depth correction matrix to enhance depth accuracy. By utilizing Simulated Annealing optimization alongside our custom calibration object, we achieve superior results compared to conventional optimization techniques. Our obtained results show that the proposed depth correction method can reduce the reprojection error from 3.12 to 2.89 pixels. Furthermore, despite the simplicity of our reconstruction method, we observe around a 22% improvement in surface reconstruction compared to factory calibration parameters. Our findings underscore the practicality and efficacy of our proposed system, paving the way for enhanced 3D surface reconstruction for real-world surface capture.

Keywords

• Azure Kinect
• depth correction
• RGB-D calibration
• sensors
• surface reconstruction