i would use opencv. there are a few blogs and forum posts around the net about feeding the live video stream into opencv. they all use the old IplImage and not the awesome cv::Mat so you may have to adjust them. but the point is that info on this workaround exists online.

-checkerboard detection of 20 frames or so
-find sub pixels (probably a large number of iterations is needed due to the crazy fish eye warp) of the inner corners of the checkerboard for each captured frame
-feed the data into cameraCalibrate
-save the intrinsic matrix and distortion coeffs in a yml file and done
you may have to recalibrate after every crash though.
the lens location may shift microscopically and mess up your principal point considering the precision of the measurement.
same goes for the glass in the lens, also the fact that this is a 92 degree fisheye doesn't help with the warp. your k1 k2 k3 p1 p2 parameters may change a lot after every crash too.