Self-calibration of asynchronized camera networks

Abstract

Calibration of camera networks is a well studied problem. However, most previous attempts assume all the cameras in the network to be synchronized which is especially difficult over large distances. In this paper we present a simple method to fully calibrate the asynchronized cameras of differing frame rates from the acquired video content directly.The presented methods either utilize content based tracked features or alternatively a light marker together with epipolar or homography based constraints to estimate the synchronization as well as intrinsic and extrinsic camera parameters. We assume two cameras within the network to be pre-calibrated (intrinsics only) using standard approaches. We validate our method with numerous simulations for noise analysis as well as real experiments. Furthermore we show how our approach can be used for robust 3D reconstruction in spite of using asynchronized cameras.

Global calibration of non-overlapping cameras: State of the art

“…Optical mirrors [34,35], [43], [47], [49], [51] [39] [ 36], [40], [48] [ 37], [45,46], [50], [52] Structure from motion [55], [57], [59][60][61] [ 53,54], [63], [66,67], [69][70][71] [56], [58], [62], [64,65], [68], [72,73] Laser projection [77,78], [80] [74-76] Close range photogrammetry [87] [ [88][89][90] range of applications in visual tracking, pose estimation and robot navigation. The method based on laser projection can theoretically work in a large range, but it is difficult to obtain high calibration accuracy because the image quality of the spot or the light plane decays as the work distance increases.…”

“…Laser projection method has a larger measurement range and is less affected by light, so often used in the global calibration method [74][75][76][77][78][79][80]. As the laser pointer is small, portable, distance measurement accuracy, methods based on the laser projection are suitable for non-overlapping field of view of the calibration parameters.…”

On the Two-View Geometry of Unsynchronized Cameras

“…The algorithm is initialized by the 7pt algorithm [14] and a zero time shift. Work [20] extended this approach by estimating difference in frame rate and using splines instead of lines. Both the above works achieve good results only when given a good initialization, e.g., on sequences less than 0.5 seconds time shift and with no gross matching errors.…”

“…The closest alternatives to our approach are the leastsquares based algorithms presented in [21] and [20]. Both optimize F or H and β starting from an initial estimate of β = 0 and T .…”

“…Both optimize F or H and β starting from an initial estimate of β = 0 and T . Method [21] uses linear interpolation from the next sample, whereas method [20] uses spline interpolation of the image trajectory and we will refer to these methods as T β -lin and T β -spl respectively. In our implementation of those methods, we used Matlab's lsqnonlin function with Levenberg-Marquardt algorithm, all stopping criteria set to epsilon and maximum number of 100 iterations.…”