Precision analysis model and experimentation of vision reconstruction with two cameras and 3D orientation reference

Rong Chen, Fang Chen, Guan Xu, Xiaotao Li, Hui Shen, Jing YuanScientific 2021 · DOI

Abstract

Active vision reconstruction is widely used in industrial manufacturing and three-dimensional inspection. The reconstruction accuracy is an important problem to be investigated for the inspection process. This paper conducts an analysis study of the reconstruction error for a vision reconstruction system utilizing a planar laser, two cameras, and a 3D orientation board. The variation principles of the spatial coordinates caused by the variations of extrinsic parameters, intrinsic parameters of the internal camera, and image coordinate points are modeled and analyzed. The analysis is validated through verification experiments, demonstrating the application potential for active-vision-based reconstructions.

Introduction

3D reconstruction is a cornerstone of modern robotics, essential for tasks ranging from autonomous navigation to high-precision industrial inspection. While binocular vision is effective for textured surfaces, it often fails on featureless or "lubricous" surfaces like polished metal. Active vision systems, which project a laser pattern onto the object, overcome this limitation.

To expand the workspace, researchers have developed "chained" systems where a mobile 3D orientation board houses an internal camera and laser, all tracked by a fixed external camera. This multi-stage setup introduces complex error propagation paths. If the external camera is off by a millimeter, or the internal camera’s focal length is slightly miscalibrated, how much does the final 3D measurement suffer? This paper provides the mathematical "precision analysis model" to answer that question.

System Architecture and Methodology

The system is modeled as a series of coordinate transformations across three primary spaces: the internal camera space, the orientation board space, and the external camera space. The laser projects a line onto the target, which is captured as a 2D image point by the internal camera. This point is then projected into 3D relative to the mobile board, and finally, the board’s position is transformed into the global coordinate system of the external camera.

Schematic of the chained vision reconstruction system showing the spatial relationship between the external camera, orientation board, internal camera, and laser plane.

Precision Analysis Model

The core innovation is the use of partial derivatives to evaluate the "variation ratio" of the final 3D coordinates. By calculating the derivative of the reconstructed point with respect to every system parameter—translation vectors, intrinsic camera parameters (focal length, principal point, skew), and image coordinates—the researchers created a sensitivity map of the entire system.

Flowchart of the reconstruction process and the derivation of partial derivatives for sensitivity analysis.

This sensitivity mapping reveals critical insights into system stability:

  • Linearity of Translation: Errors in the translation vector (the physical distance between components) have a linear relationship with the final reconstruction error, making them predictable and easier to compensate for.
  • Sensitivity of Image Coordinates: The system is most sensitive to noise in the 2D image coordinates ($x, y$ pixels). Small amounts of sensor noise or sub-pixel extraction errors lead to non-linear, significant jumps in the 3D output.
  • Chained Error Propagation: Because the system is "chained," an error in the external camera's perception of the orientation board compounds every error occurring "upstream" in the internal camera.

Experimental Results

The model was validated through both computer simulations and physical experiments using a custom-built setup involving a 1280x960 resolution camera and a sub-pixel accurate laser projection system. The primary benchmark was the reconstruction of laser points on a high-precision chessboard pattern at distances ranging from 600mm to 1300mm.

Experimental results perfectly matched the model’s prediction that reconstruction coordinates increase linearly with the growth of the distance between the external camera and the orientation board. Furthermore, simulation plots successfully mapped the "error surfaces" for intrinsic parameters, showing that the $Y$-axis coordinate is generally more sensitive to focal length variations than the $X$ or $Z$ axes in this specific configuration.

Conclusion and Future Work

This study provides a rigorous mathematical framework for building more precise industrial inspection robots. By quantifying exactly how small fluctuations in calibration or image noise translate into spatial inaccuracies, it offers a roadmap for system optimization.

Future work will aim to address the complexity of rotation matrix errors (angular "tilt") and environmental factors, such as heat-induced lens distortion or laser speckle, which may introduce biased errors in real-world industrial settings.

Summary updated

Abstract

Active vision reconstruction is widely used in industrial manufacturing and three-dimensional inspection. The reconstruction accuracy is an important problem to be investigated for the inspection process. The paper conducts an analysis study of the reconstruction error for the vision reconstruction with a planar laser, two cameras and a 3D orientation board. The variation principles of the spatial coordinates caused by the variations of the extrinsic parameters of the cameras, intrinsic parameters of the internal camera, and image coordinate points of the internal camera, are modeled and analyzed in this paper. The analysis is also proved by the verification experiments, which provides the application potential for other active-vision-based reconstructions.
BibTeX
@inproceedings{chen2021precision,
  author = {Rong Chen and Fang Chen and Guan Xu and Xiaotao Li and Hui Shen and Jing Yuan},
  title = {{Precision analysis model and experimentation of vision reconstruction with two cameras and 3D orientation reference}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-83390-y},
}
Figure