This paper suggested a target segmentation scheme based on morphological processing and texture analysis. Firstly

an opening-closing filter bank performed on the gradient image for watershed segmentation is proposed. The filter bank concerns with eight structure elements

each is selected to represent the allowed neighbor patterns and used to generate a reference image for segmentation. This reference image is more suitable than the gradient one when calculating the watersheds

because its main structures are preserved while pseudo-contours are discarded. Then binary marker extraction is carried out to get initial partition. These region markers mark the relevant objects and backgrounds to act as the minima set of segmentation

the initial object boundaries are finally obtained by computing the watersheds. Secondly

a criterion with respect to the homogeneity and the contrast of textures is defined to merging regions

it weighs the contrast of two neighbor patterns and adapts itself to different sizes of neighbor patterns. Experimnts on several video sequences are presented to demonstrate that the morphological filter bank could prevent the result from over-segmentation and the criterion is also validity.