Realtime VDO stabilizer for small UAVs using a modified homography method

Abstract

An approach to video stabilization for small unmanned aerial vehicles (UAVs) using computer vision techniques is proposed. As gimbals (mechanical video camera stabilizer) cannot be installed on UAVs with small payload, captured videos can be shaky. A development of `software image stabilizer' has been suggested by a military institute to alleviate such adverse effect. We, similarly to other researchers, choose homography theory for warping the shaky video. The novelty is we can cope with large rotational difference. This is possible by revamping homography matrix, using information from an orientation sensor attached to a video camera to acquire the camera's orientation in real time. Then, extrinsic parameters derived from the sensor and pre-computed intrinsic parameters were used to generate our modified homography matrix. Furthermore, calibration between camera and sensor is necessary because an alignment of sensor and camera is imperfect. The calibration relies on an Iterative Least Square method.