Depth Data based Chroma Keying using Grab-cut Segmentation

The research presents a depth-image based automatic object segmentation for chroma key editing in multimedia applications. Depth data taken from advanced depth data capturing devices like Microsoft Kinect has a key role in this research. The proposed approach uses both color and depth data and this hybrid segmentation generates results with clear foreground object boundaries. The present system is exclusively designed for segmenting human subjects from a chroma keyed scene. An Aggregate Channel Feature (ACF) based human detection is also employed here to eliminate false detection due to other foreground objects. The depth data results in dark regions may create small errors over edge pixel segmentation. So, the whole process is carried through a sequence of image processing techniques. The pixels nearby the head portion first restored using K-means clustering and then a coarse level segmentation of the human subjects is obtained using Fuzzy C means segmentation. A color characteristic-based segmentation us used here to eliminate most of the background pixels from the foreground subjects. After this coarse level segmentation, an adaptive Tri-map generation is employed and the ultimate fine level segmentation is achieved using Grab-cut Segmentation to generate the foreground human subjects with accurate edge boundaries for matting chroma keyed images/frames. Experimental results validate the segmentation results and its ability for an error free automated segmentation in editing chroma keyed images or video.