Abstract
In this thesis, we present a novel stabilization system for the videos recorded by dashcams. Our system takes the effects of optical distortion into consideration and attempt to estimate the shaky trajectories under radial distortion. Thus, we can obtain stable videos by removing the jitters and distortion effects. Video stabilization is a technique to enhance video quality by removing vibration in the videos. In dashcam videos, the optical distortion often interferences the estimation of the camera motions. For the input video, we first estimate the distortion parameters from the video. We estimate the distortion parameters from point correspondences computed from multiple frames in the video. The division model is applied for distortion correction in this work. Then, we estimate the optical flow from consecutive frames in the corrected video. The ego-motion of the camera can be further estimated by fitting the estimated optical flow vectors to the 3D motion model. Then, a filtering procedure, moving average, is applied to the camera motion parameters to obtain a smooth camera trajectory. We use the smoothed parameters to synthesize the stabilized video. In order to justify our stabilization system, we demonstrate the proposed algorithm on synthetic and real-world videos. We also compute the inter-frame fidelity as a quantitative metric for evaluating the video stabilization result.