Contents I Intruduction
For mobile robot visual information can more actually reflect external environment than other informations. In order to acquire more visual information color image is usually adopted in visual systems of mobile robot. For tasks of object recognition, tracking, self-localization and navigation, segmentation of visual color image is necessary, whose result is key to the following processing. In the field of color image segmentation extensive research has been conducted. Major categories of existing segmentation approaches included: (1) feature-based segmentation (e.g., color clustering, normalized cut and mean shift), (2) edge-based segmentation (e.g., edge flow, snake, balloon), (3) region-based segmentation (e.g., split-n-merge, region growing), and (4) hybrid segmentation (e.g., region competition, EDISON)[1],[2],[3] Among these methods clustering segmentation is preferable, whose basic idea is that color image segmentation is regarded as a problem of dataset clustering and image is segmented by methods of dataset clustering according to the rule of colors' similarity in 3D color space. Fuzzy c-means, self-organizing map and competitive learning are widely employed in clustering segmentation [3], [4], which belong to supervised learning and need specify the number and centers of clusters in advance. If not, result of segmentation may become considerably worse. But for visual images' segmentation of mobile robot its captured images change consecutively during it moving, and it is a difficult task to specify the number and centers before segmentation. To solve this problem researchers have proposed many different methods.
