TL;DR
This paper introduces a novel and convenient camera calibration method that estimates both the extrinsic camera parameters (position and orientation) and the focal length using a single image of two arbitrary coplanar circles. This approach simplifies calibration by allowing for occluded circles and easy-to-create calibration objects, addressing a significant need in real-world vision applications.
What Problem Does This Paper Address?
Accurate camera calibration is crucial for many computer vision tasks, but it often requires complex setups, specially designed calibration objects, and multiple images. In real-world applications, camera settings, including zoom, are frequently adjusted, necessitating recalibration. Existing methods often struggle with ease of use, requiring precise point correspondences, multi-view images, or specific calibration patterns. This paper aims to provide a simpler, more flexible calibration method that can be performed with a single image and a readily available calibration object, while simultaneously estimating the focal length, which is often altered with zoom adjustments.
Methods
The core of the method relies on the geometric properties of how circles project onto an image plane. When a circle is viewed with perspective projection, it appears as an ellipse. The paper first establishes the mathematical relationship between an ellipse in an image and the oblique elliptical cone formed by projecting lines from the camera's optical center through the ellipse. This cone is related to the original circular cone.
The key innovation lies in using two coplanar circles. If the focal length is known, the unit normal vector of the plane containing the circles can be estimated independently from each detected ellipse. However, if the focal length is unknown, these estimated normal vectors will not align. The method leverages this discrepancy: by minimizing the difference between the two estimated normal vectors (specifically, by minimizing $(N_1(f) \cdot N_2(f) - 1)^2$), the correct focal length ($f$) and the true unit normal vector ($N$) of the supporting plane can be simultaneously determined. This process also resolves ambiguities in determining the circle centers. Once the focal length and the plane's orientation are known, the extrinsic camera parameters (rotation and translation) can be derived.
Datasets
The paper describes experiments using two types of datasets:
- Simulated Images: Computer graphics were used to synthesize images of scenes with coplanar circles. Two "camera settings" (case-1 and case-2) were defined with different resolutions, focal lengths, tilt angles, roll angles, and distances to the supporting plane. 32 images were synthesized for case-1 and 17 for case-2, each containing two randomly selected circles from a larger set.
- Real Images: The method was tested on real-world images featuring coplanar circle objects such as CD discs, manholes on roads, and widening rings on water. These images were captured using digital still and video cameras with zoom lenses.
Key Insights and Ideas
- Simplicity from Geometry: The method cleverly exploits the geometric relationship between projected circles (ellipses) and cones to derive calibration parameters.
- Simultaneous Focal Length and Extrinsic Calibration: A significant advantage is the ability to estimate both focal length and extrinsic parameters from a single image, which is crucial for cameras with adjustable zoom.
- Robustness to Occlusion: The method does not require the entire circles or their centers to be visible, making it more practical for real-world scenarios where objects might be partially occluded.
- Easy-to-Create Calibration Object: The use of arbitrary coplanar circles means the calibration pattern can be easily created, even by drawing on the ground, reducing the need for specialized equipment.
- Single-View Advantage: Performing calibration with just one image significantly simplifies the process compared to multi-view methods.
Key Results
The experimental results, summarized in Table 1 for simulated data and Table 2 for real images, demonstrate the effectiveness of the proposed method.
- Simulated Data (Table 1): The estimated parameters (focal length, roll angle $\beta$, and tilt angle $\theta$) showed low RMS errors and standard deviations compared to the ground truth values used for synthesis. For "case-1", RMS errors were 5.52 pixels for focal length, 0.36 degrees for $\beta$, and 0.57 degrees for $\theta$. "Case-2" showed slightly higher errors, with RMS errors of 7.19 pixels for focal length, 0.11 degrees for $\beta$, and 0.51 degrees for $\theta$.
- Real Images (Table 2): The method successfully estimated focal lengths and unit normal vectors for various real-world scenes. For example, for "Table scene 1", the estimated focal length was 901.0 pixels, and the unit normal vector was (0.03, 0.83, 0.56). For "Manhole on road", the focal length was 4164 pixels with a normal vector of (0.05, 0.97, 0.26). The paper also shows synthesized "vertical views" (Figure 3) which appear geometrically consistent, suggesting accurate parameter estimation.
Limitations and Future Work
- Intrinsic Parameters: The method assumes that intrinsic camera parameters, except for the focal length, are already calibrated. The paper notes that uncalibrated intrinsic parameters (other than focal length) and radial distortion were reasons why converted images of real scenes did not always show perfect circles.
- Future Work: The authors plan to extend their method to estimate the image center and radial distortion parameters in future work.
Summary updated