New Intrinsic Calibration Procedure

According to a ROS users survey that was conducted in 2014, the most popular hardware to integrate with ROS is a camera. Cameras are often used to perceive the environment or to localize robots and are a critical component of the sense-plan-act capability that ROS enables. Over the past two years, the ROS-I team has been working to create an industrial calibration library that supports both intrinsic and extrinsic calibration of vision sensors. What is novel about the library is that is can handle groups of heterogeneous sensors that may be static, or mounted to a robot, or some combination thereof. And it coordinates with MoveIt! to automate calibration procedures in which robot motion is required during calibration. While the extrinsic calibration routines are well in hand, the intrinsic calibration algorithm, which is based on a popular lens distortion model, resulted in higher parameter variance than was expected based on residual errors. This is particularly true for the focal length parameter, which is essential for correctly interpreting the size of objects in the scene. The ROS-I team has developed a novel camera intrinsic calibration technique that is both computationally faster and provides superior results to the methods commonly employed in machine vision.

The optimization procedure outlined by Zhang and automated by both OpenCV/ROS, and Matlab orchestrate the collection of a set of images of a calibration target. Both the extrinsic pose of the camera and the intrinsic parameters themselves are determined by minimizing the re-projection error. Using these methods, the residual re-projection error is on the order of ¼ pixel/observation or less. However, the variances of focal length and optical center are much higher, typically being 20 pixels and 5 pixels respectively. This is due to correlation between parameters of the distortion model with the focal length parameter.

The new procedure developed by the ROS-I team reduces parameter variance to be on par with the residual error. It requires only 10 to 20 images, but each is taken a known distance apart with little or no skew (refer to images). The new procedure estimates the extrinsic pose for the first image, and constrains the optimization to use the known pose relationship for subsequent images. The focal length and optical center are significantly better constrained. Using the resulting intrinsic calibration parameters for a given camera yields significantly better extrinsic calibration or pose estimation accuracy. Try out the Intrinsic Camera Calibration (ICC) tutorial that is posted on the ROS-I wiki.

The new intrinsic calibration procedure requires one to move the camera to known positions along an axis that is approximately normal to  the calibration target.