Research On Image Matting Algorithm Based On Kinect

Foreground extraction is a hot topic in the research of image processing. With thecontinuous improvement of theory, the foreground extraction technique has developedfrom image segmentation to image matting based on high level alpha value computation.So far, image matting has already witnessed applications in realms of film andadvertising production, industry and agriculture, environmental monitoring, medical andhealth care and so on. Image matting is an ill-posed problem，and many natural mattingalgorithms often require a user provide the segmentation to identify background,foreground, and unknown regions. This segmentation is called a trimap. The algorithmsthen use information given by the trimap to disambiguate the unknown regions. Oneproblem with trimaps is that they must generally be drawn by hand, either for eachframe or at key frames. This makes the algorithms difficult to extend to video becausemanually creating trimaps for many frames is far too timely a task when dealing withlong sequences. In addition, most natural matting algorithms work only on the imagedomain and don’t take depth information into consideration.With the research and development of depth camera, there have been a series ofmatting algorithms based on depth maps in recent years. Each pixel in depth maprepresents the distance between one point in space and depth camera, which isconvenient for telling foreground area from the background area. So far, depth camerasare mainly divided into two categories: Time-of-flight camera and Kinect.In this paper, we studied the traditional matting algorithms based on TOF cameraand analyzed their merits and demerits. With depth maps (or user maps) generated byKinect, we can obtain a matte of Kinect color map. This paper introduces a newimproved joint bilateral filter, a novel trimap automatically generation method andacquire a matte combing the shared matting algorithm. Analyzing the Geometricrelationship between HD-camera and Kinect, we can also acquire a matte of HD-cameracolor image. Experimental results demonstrate the robustness and real-time performanceof our approach.