针对视频稳像中实时任务的需求特点，提出一种基于单应轨迹的视频稳像算法。估计序列图像帧间的单应变换，并将该变换作用于图像窗口的4个角点，从而对每帧图像产生4个短的单应轨迹以代表视频短时间内的运动。利用关联卡尔曼滤波器以一种连续方式对不同帧的单应轨迹进行平滑。对图像合成采用包括性和相似约束以提高结果视频的可观性质量。该算法以在线方式工作，消除了缓存输入图像帧导致的延迟，具备不依赖于复杂的3D重建和长距离特征跟踪的优点，并有效避免了单应模型表达视频运动模型的误差积累问题。实验表明该算法能够有效对包含2D和较复杂3D场景的视频进行稳像，并且能够达到实时处理速度。%Motivated by the demands of real-time video stabilization,a real-time video stabilization algorithm based on homography trajectory is proposed.For each input frame,our approach regenerates four short homography trajectories by applying inter-frame homography transformations to the four corners of image rectangle.An associate Kalman filter is then performed to smooth these transformational trajectories.Finally,at the stage of image composition,constraints of inclusion and similarity are considered for selecting a visually plausible stabilized video.The proposed method can offer real-time video stabilization and it can remove the delays for caching coming images.In addition,our method does not rely on sophisticated 3D reconstruction or long-range feature tracking and it can effectively relieve the errors introduced by using homography to model the video motion. Experiments show that our approach can offer real-time stabilizing for videos with 2D scenes or the 3D scenes with moderate depth variation.
展开▼
