Research On Foreground Segmentation Based On Color And Depth

Recently, with the rapid development of stereo vision,3DTV, intelligent transportation, as a key technology in these applications, foreground segmentation of image has attracted more and more attention of researchers in computer vision. As the parts of image which can represent the image content, foreground objects are the most concerned part in the high-level applications such as image analysis, image understanding.There are a variety of foreground segmentation methods, but most of the methods have disadvantages in practical application, such as the semantic meaning cannot be guaranteed, and boundaries are not accurate. In view of the widespread applications of foreground segmentation in the industry and research, this thesis proposes a foreground segmentation method based on the depth map, aiming to solve the problem of uncertainty in foreground segmentation.The major content of this thesis is to study foreground segmentation based on blending depth and color information. In image pairs consist of depth map and color image, semantic foreground area is given on the basis of a color image segmentation combing with the depth map grouping. How to choose the color space and color image segmentation algorithm to segment the color image, how to repair the original depth map of low quality captured by depth camera Kinect, how to determine the foreground area and overcome the over-segmentation of color image are main problems to be solved in this thesis. We firstly introduce the concept and difficulties of image segmentation, especially detail the difficulties of foreground segmentation. Then we describe color image segmentation algorithm, expounds the principles and application of Mean shift for color image segmentation. We also introduce the depth map in stereo vision, describe the methods using the stereo matching to calculate the depth map, as well as the direct use of depth camera to obtain depth map. Then we are focused on foreground extraction method with depth and color information, present an algorithm to merge segmented regions of color image on the basis of depth grouping, and complete experiments of extracting single and multiple foreground objects from different images. Finally we summarize our work, and point out the problems and defects existing in the method, specify the direction to improve the method.In this thesis, we develop an intuitive and novel foreground extraction method by blending of color and depth information on the basis of the stereo vision, our method is validated through experiments, which show our method can accurately segment the foreground area in the image. The proposed method is applicable in many fields.