Automatic Camera Calibration Applied to Medical Endoscopy

We present a calibration algorithm for cameras with lens distortion, that uses a single image of a planar chessboard pattern acquired in general position. The radial distortion is modeled using the first order division model, and the method provides a closed form estimation of the intrinsic parameters and distortion coefficient. The experimental evaluation shows that the calibration accuracy is comparable to state-ofthe-art algorithms requiring multiple input images. We believe that our approach is particularly well suited for the the calibration of medical endoscopes in computer aided surgery. Since the lens is mounted on the camera before each usage in the OR, the calibration procedure must be performed by the clinical practitioner with minimum effort. We solve this problem by proposing a fully automatic procedure that requires no human intervention other than acquiring the calibration image1. Camera calibration can hardly be avoided in the context of imagebased computer aided surgery. However, developing an effective method for geometric calibration of a medical endoscope is a very challenging task [4]. The reasons are the following: (i) the calibration must be very accurate; (ii) the endoscope optics introduces strong radial distortion that must be considered in the projection model; (iii) the calibration procedure has to be performed by a non-expert in the Operation Room (OR), which requires the method to be robust and fully automatic. Geometric camera calibration is a well studied topic, for which there are several methods and software currently available. Bouguet’s implementation of Zhang’s method for calibrating a camera from a minimum of three grid images is particularly popular [2, 5]. Unfortunately the calibration procedure does not meet the usability specifications stated above. The approach requires the acquisition of several grid images and the manual selection of corner points. This selection, although easily automatized for standard perspective images, it is very problematic in the presence of strong lens distortion. We can hardly imagine a medical doctor being eager to acquire 10 to 20 grid images during his regular clinical practice, and then struggle to manually select points over all this imagery. Wengert et 1Patent Pending (number 20091000069650) 0 2 4 6 8 10 12 290 300 310 320 330 340 350 360 370 380 390 Image # C en te r C oo rd in at es ( pi xe l) Estimation of the Center / Principal Points c x − Calib from Single Image