Robust and Accurate Calibration Point Extraction with Multi-scale Chess-Board Feature Detector - Advances in Image and Graphics Technologies - page 155

This scheme performed badly when the grid is distorted because of optical distortion or
on a non-planar surface.
Yu and Peng[7] adopt a pattern-match method to find specific features by measuring
the correlation over all the image. This method fails when the grid is rotated relative to
the patterns in store.
Sun et al. [8]describe a method which places a rectangular or circular window over
every position of the image before achieving a 1D binarized vector along the perimeter.
The positions where the vector has four regions are determined to be chess-board
vertices. This method is somehow rather slow and prone to noises. The performance
also relies on the result of binarization.
Shen Cai and Zhuping Zang [9] designed a deformed chessboard pattern for auto-
matic camera calibration, but the precision and robustness of their method needs to be
improved.
Stuart Bennett and Joan Lasenby [10] offer an instructive approach where they
emphasize the importance of chess-board detector and design ChESS which is both
simple and quick. It is rather accurate but tends to produces much more false features.
The sing-scale property also aggravates its limitation.
Besides, some open source labs such as OpenCV [11] and Matlab [12] tools are
widely used for chess-board grid detection for their convenience, but they both have
some problems concerning to accuracy and robustness.

3 Corner Detection Algorithm

Our chess-board marker detection algorithm for camera calibration mainly consists of
two steps. First, we obtain all the chess-board features with the help of the multi-scale
chess-board feature detector and eliminate outliers from them. Then we organize the
vertices found in former step and map them to 3D locations in the marker. The detailed
algorithm will be introduced by these two steps.

3.1 Detecting Chess-Board Vertices

In this section, we propose and justify several properties of the chess-board calibration
pattern , based on which we design a single-scale annular chess-board feature detector
and then develop it into a multi-scale detector to ensure the robustness and precision of
the detection.

3 . 1.1 Annular Chess-Board Detector
When we put a sample circular window over a chess-board vertex, the points on op-
posite sides of a diameter tend to have similar intensity and those on perpendicular
radii should have very different intensity ( Fig. 3 ). Based on this observation, we build
an annular chess-board detector which has two kinds of energies: and .
Unlike [10] where the two properties are combined into one, we find it produces
much fewer false features if considering them separately.