Realtime Object Segmentation Using Background Information

• Main Authors: Po-Wen Cheng, 鄭博文
• Other Authors: Chia-Wen Lin
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/88925437382624784113

碩士 === 國立中正大學 === 資訊工程研究所 === 91 === This thesis addresses two topics of real-time video object extraction with a still background: one is about object segmentation using prestored background information, while the other is about object extraction without prestored background information. For both scenarios, we propose robust video object extraction algorithms using background subtraction. For the application scenario with prestored background information, our proposed framework consists of three main procedures: preprocessing, object segmentation, and postprocessing. In the preprocessing stage, the background images are captured and analyzed to extract statistics for use in the following segmentation procedure prior to objects coming into the scene. The segmentation algorithm performs background subtraction and then computes and combines two statistical features, namely, normalized statistics and high-order statistics, from the background subtracted images to separate moving objects from the background. After the initial segmentation, a series of effective postprocessing techniques, including shadow removal, region growing, nonlinear filtering, and boundary watershed, are then put forth to refine the segmentation mask. The proposed method can produce satisfying results with pixel-wise precision. Furthermore, it is robust to the effects of camera noise, the changing of lighting condition, and the effects of shadow. The proposed method is useful in applications with a still background which can be captured and analyzed beforehand, such as virtual conferencing and video surveillance. For those applications that the background images cannot be captured and analyzed beforehand, we propose a background registration technique to construct the background information, with the assumption that the segmentation mask of the first frame can be accurately obtained through the assistance of human interaction. In the proposed background registration scheme, only values of those pixels considered as reliable background justified with a confidence measure are registered into the background memory. With the registered reliable background information, background subtraction is then performed to extract the foreground pixels, and then followed by a postprocessing procedure to refine the segmentation results. In this method, in order to resolve the problem of false segmentation due to uncovered background, we propose efficient schemes which extract and exploit moving object edges to effectively identify the uncover background regions. Good segmentation performance is demonstrated by the simulation results.