The Automatic Board Selection Algorithm

The goal of the RADLOCC algorithm is the automatic extraction of calibration board lines from 2D laser scans. Extrinsic laser-camera calibration based on observing a well defined object requires the extraction of that object from the laser scans (this is also true for the camera). Rather than relying manual extraction, the RADLOCC algorithm was proposed to achieve the automatic extraction of chessboard calibration boards – which are the calibration object of choice. This page briefly mentions key ideas in the algorithm; for a detailed explanation, please refer to [1].

The proposed straight line criteria is that a set of consecutive points in the laser scan is considered a straight line if the maximum distance from any point of that set to the straight line formed by joining the end points is less than a certain threshold. The figure below visualises the criteria. The red line is the line formed by joining the end points. The black lines represent the distances to be tested. The maximum distance should be less than a certain threshold. In the RADLOCC algorithm, two thresholds are used for two different purposes. When finding the initial estimate, a loose threshold of 5cm is used. This is necessary to include as many board lines as possible, even if they are noisy or include some outlier points from the frame of the calibration board or persons around. For the following stage however, after the estimate has been optimised, a tighter threshold is used. This ensures the removal of outlier points and noisy board scans.

One of the metrics used for the classification process is the extent of match to the initial estimate of the transformation. In other words, assuming an initial estimate of the transformation exists, each laser scan is transformed to the camera plane and the distance of each point in the scan from the calibration board is measured. Depending on the accuracy of the initial estimate, the score will vary in its ability to signify the calibration line. Therefore, a refinement loop is applied to improve the estimate of the transformation. Initially, data points are selected based on frequency and length metrics only. These data points are used to obtain a rough initial estimate, which is then incorporated in the classification process which may choose a different set of data points. This is repeated until the data points selected do not change. Once convergence is reached, the data points are used for the final optimisation of the calibration using the last initial estimate obtained.

After straight lines are extracted they are subject to a classification process which selects the correct calibration board line. The metrics used for the classification process are:

1. The initial estimate (if available)
2. The frequency of the point range.
3. The length of the lines.