A novel online learning based video segmentation algorithm was proposed, combining both the global and local information of video images. The videos were pre-segmented by the unsupervised image segmentation method firstly, and then the coarse foreground was extracted by the detection of the classifier. After that, the final optimal pixel-wise segmentation was achieved by using spatial-temporal Conditional Random Fields, and the classifier was updated with the constraints of the segmentation result. Meanwhile, a balance sampling strategy and a sample-updating approach supervised by segmentation were proposed, to improve the accuracy and stability of the classifier on initialization and updating separately. Experiments on challenging video sequences show that the proposed method highly improves the precision and the stability of video segmentation with low time cost, compared to state-of-the-art methods.
online learningunsupervised image segmentationincremental learning classifierconditional random fieldsvideo segmentation
NETL
NSTL
万方数据
维普
