An Unsupervised Video Stabilization Algorithm Based on Key Point Detection

In recent years, video stabilization has improved significantly in simple scenes, but is not as effective as it could be in complex scenes. In this study, we built an unsupervised video stabilization model. In order to improve the accurate distribution of key points in the full frame, a DNN-based key-point detector was introduced to generate rich key points and optimize the key points and the optical flow in the largest area of the untextured region. Furthermore, for complex scenes with moving foreground targets, we used a foreground and background separation-based approach to obtain unstable motion trajectories, which were then smoothed. For the generated frames, adaptive cropping was conducted to completely remove the black edges while maintaining the maximum detail of the original frame. The results of public benchmark tests showed that this method resulted in less visual distortion than current state-of-the-art video stabilization methods, while retaining greater detail in the original stable frames and completely removing black edges. It also outperformed current stabilization models in terms of both quantitative and operational speed.