Segmentation Of Moving Objects In Video Sequences

As a basis of video processing, moving objects segmentation is widely used in military technology, vision navigation, pattern recognition, intelligent monitoring, video encoding and other fields. Based on the study of some commonly used segmentation methods, this paper proposes a new segmentation method of moving objects fusing the background subtraction and multi-scale edge images in video sequence with static background.In order to get correct object region by background subtraction algorithm, the initialization and update methods of background has been researched to aim at solving the problem that the process of background subtraction is highly sensitive to brightness change. First of all, by using the foreground separation, the initial background has been constructed based on the improved K-Means clustering algorithm; To minimize the detection error which caused by the brightness difference of scene during a short time in the process of following background construction, a mechanism of brightness compensation in the background has been proposed; At the same time, to adapt the slight change of background and the change of the background structure, a short time and long time mechanism of updating background to refresh the background information have been adopted respectively. After obtaining effective background, initial object template will be obtained through background subtraction.Due to the noise and the impact of contrast between moving objects and background, the template of moving objects by background subtraction are often incomplete. Especially the accuracy of marginal parts is comparatively low. In order to obtain more precise outline of objects, the method of combining edge information of multiple scales is proposed to obtain comparatively complete object outlines based on the advantage of multi-scale images, which made up for a deficiency of background subtraction algorithm. Focusing on the impact of shade during moving object segmentation, the shade detection has been run to get rid of the impact of object shade by using the difference between shade and background in the HSV space. Finally, the precise moving objects have been segmented through post-processing.