Real-Time Digital Image Stabilization Algorithm using Sage-Husa Filter

We propose a digital image stabilization algorithm based on improved SageHusa filter. Global motion estimation is obtained by detecting interest points from reference frame, matching points in current frame, and solving motion equations. The motion includes intentional camera scan and unwanted jitter; therefore, we need to smooth global motion sequence to remove jitter. The Sage-Husa filter is improved by modifying process and observation noise using their statistical property. The prediction error matrix can be adjusted according to the property of innovation sequence to avoid filter divergence. Finally each current frame is warped by fast motion compensation in consideration of the linear storage of the image. We investigate the filter accuracy and visual quality by experimental evaluation. The results show that the proposed algorithm is capable of smoothing inter-frame jitter and tracking real scene at real time speed.