Computer vision on rolling shutter cameras
(Linkoping University, Sweden)
The majority of cameras sold today have CMOS sensors with electronic
rolling shutters. In the absence of a mechanical shutter this means that
instead of a global frame exposure, frames are exposed in a line-by-line
fashion, where the lag between the first and last row readout (the
readout-time) is measured in tens of milliseconds. A rolling shutter
causes geometric distortions, whenever either the camera, or objects in
the scene are moving, and this can have severe implications for any
computer vision algorithm that is geometry based.
In this talk I will describe our efforts to design novel algorithms
that take the rolling shutter into account. I will cover versions of
video rectification, video stabilization, and bundle adjustment of
video. Finally I will show some recent results on how to improve 3D
mapping using the Microsoft Kinect sensor, by taking its rolling
shutter into account.