Video Object Segmentation Based On Tracking

Video object segmentation, as a significant component in computer vision, is to extract the target of interest. It has been largely researched and widely applied in many areas, such as intelligent transportation, pose recognition, video surveillance and video compression.Given the facts that the unknown object tracking has been well improved, we attempt to deal with the problem of object disappearance and reappearance in video object segmentation through integration of tracking. For the first frame of a video, it is segmented by the classical Grab Cut, and its result is treated as the initial input of our segmentation model while the interactive box is used as the input of tracking algorithm. For the rest frames of the video, we first track the objective target, the current frame is segmented after obtaining tracking box, and then the segmentation result is utilized to update segmentation model by P-N learning method. The performance of tracking algorithm can dramatically affect our segmentation result. Our main contributions are as follows:1) We review the latest methods of tracking and segmentation respectively. The P-N learning, random fern based on BRIEF(Binary Robust Independent Elementary Features) and Tracking-Learning-Detection(TLD) is mainly researched.2) To get a real-time tracking algorithm, we re-design the detector of TLD and propose a new long-term tracking algorithm TLLD(Tracking-Learning-Local-Detection). Considering that full screen detection is time-consuming, an approach that adaptively generates the detective range for visual object tracking is presented, which improves the detecting efficiency. At the same time, to alleviate the shortcoming that caused by less samples, a Random Fern classifier and the corresponding P-N learning method are introduced to increase detecting accuracy. Our experimental tracking results on the pubic TLD datasets demonstrate that our method quantitatively and qualitatively outperforms the original algorithm.3) Built on the framework of graph-cut, a BRIEF based Random Fern classifier is integrated to the common color Gaussian mixture model to obtain a more reliable probability model. And a video object segmentation method is put forward by combining the model and TLLD. Our experimental results indicate that the proposed method can effectively resolve object disappearance and reappearance in the video.