Video Cutout Using Its Illumination Property And Texture Information

As a fundamental task in Computer Vision, segmentation has beendeeply studied for many decades. With the development of technology,video segmentation has drawn attention these years and several approacheshave been proposed. The results have also been used by several othervision tasks, for example object tracking, recognition, and videoclassification. Since segmentation is an ill-posed problem, user interactionis required in order to get accurate results. To deal with the large amountof video data, an efficient algorithm is also required. For this purpose,graph-cut technique is used in this work in order to segment the foregroundobjects from the background. An inherent challenge of this technique is tointegrate different cues in an efficient way in order to get satisfied results.In this thesis, a video segmentation algorithm is proposed. Colorinformation, texture information as well as motion cues are integrated intothe framework to get space-temporal consistent results. The self-quotientimage technique is applied in order to get lighting invariant results. As oneof the basic properties of objects, color spaces, as the representation ofcolor, have been studied as well. Texton has been applied to extract textureinformation and optical flow has been used to get the motion informationfrom the video sequences. A video segmentation model is proposed as wellin order to guarantee the algorithm’s efficiency. Detailed experiments areconducted in order to study the role of each cue in the framework. Satisfiedsegmentation results are achieved as well.