A New Method for Absolute Pose Estimation with Unknown Focal Length and Radial Distortion

Kai Guo, Hu Ye, Honglin Chen, Xin GaoSensors 2022 · DOI

TL;DR

This paper introduces a novel, efficient, and robust method for estimating a camera's absolute pose (position and orientation) when its focal length and radial distortion are unknown. By cleverly decomposing the problem and leveraging a key geometric property of radial distortion, the method achieves accurate results with minimal 2D-3D point correspondences, making it particularly useful for wide-angle and zoom lenses.

What Problem Does This Paper Address?

Estimating a camera's absolute pose is fundamental to many computer vision tasks like 3D reconstruction and augmented reality. Traditional methods often assume known intrinsic camera parameters, such as focal length and distortion. However, in real-world scenarios, especially with wide-angle or zoom lenses, these parameters can be unknown and vary, significantly impacting pose estimation accuracy. Existing methods for estimating pose with unknown intrinsics typically require more 2D-3D point correspondences, making them less efficient. This paper tackles the challenge of accurately estimating camera pose with unknown focal length and radial distortion using the minimum number of 2D-3D point correspondences (three) and a known camera position.

Methods

The core innovation of this paper lies in decomposing the complex problem into two manageable sub-problems:

  1. Focal Length and Radial Distortion Estimation: The authors exploit a crucial geometric characteristic of radial distortion: the orientation of an image point relative to the center of distortion remains unchanged, even though its distance from the center is altered. This property allows them to derive a system of equations to iteratively estimate the focal length and up to four orders of radial distortion. Importantly, they use initial values derived from a model without radial distortion, which are close to the global optimum, leading to efficient and accurate convergence. This approach is flexible and can work with different radial distortion models (e.g., division and traditional models).
  2. Absolute Pose Estimation: Once the focal length and radial distortion are estimated, the image points can be "undistorted." This effectively transforms the problem into a standard absolute pose estimation scenario with known intrinsics. The paper then uses geometric linear constraints with two 2D-3D point correspondences to efficiently determine the camera's rotation and translation.

Datasets

Not discussed in the paper. The authors synthesize data for experiments and use real images for qualitative and quantitative evaluation.

Key Insights and Ideas

  • Geometric Invariance is Key: The paper's central insight is that the angular relationship of image points to the distortion center is invariant under radial distortion. This geometric property is the foundation for their novel estimation method.
  • Decomposition Simplifies Complexity: Breaking down the problem into estimating intrinsics first, then using those to estimate pose, significantly simplifies the overall solution.
  • Good Initial Guesses Matter: Using undistorted model results as initial guesses for the iterative estimation of focal length and distortion leads to faster convergence and a single, globally optimal solution, avoiding the multi-solution ambiguity of some other methods.
  • Known Camera Position is Powerful: Leveraging a known camera position (obtainable from devices like RTK or total stations) allows for a more constrained and efficient solution, requiring fewer 2D-3D correspondences.

Key Results

  • Accuracy and Stability: The proposed method demonstrates high accuracy and numerical stability for pose estimation, even with unknown focal length and radial distortion, on both synthetic and real-world data.
  • Robustness to Noise: The method shows good robustness to noise in image points and even camera position, outperforming a comparative method (Josephson's method) in several aspects.
  • Efficiency: While slightly slower (3.4%) than a general solver in their tests, the method achieves this with significantly fewer correspondences and better accuracy, especially in challenging scenarios.
  • Real-World Performance: Experiments with real images, including those with heavy distortion, show that the method can effectively undistort images and provide accurate pose estimates.

Limitations and Future Work

  • Computational Speed: The paper acknowledges that their method is slightly slower than a general solver due to the iterative step size in the LM algorithm. They suggest that adjusting this step size can balance computational speed and accuracy.
  • Adaptive Step Size: Future work will focus on developing an adaptive iteration step size for the LM algorithm to automatically optimize the trade-off between computational speed and accuracy.

Summary updated

BibTeX
@article{guo2022new,
  author = {Kai Guo and Hu Ye and Honglin Chen and Xin Gao},
  title = {{A New Method for Absolute Pose Estimation with Unknown Focal Length and Radial Distortion}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22051841},
}
Figure