Improvement of Foreground Segmentation Using Local Patch Variation

• Main Authors: Chien-Hung Wu, 吳建宏
• Other Authors: G.S. Hsu
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/55266322205216359537

碩士 === 國立臺灣科技大學 === 機械工程系 === 98 === Pixel-based dynamic background models, such as GMM or Codebook, are designed to capture the variation of a pixel’s value across time without considering the relationship between the pixel and its neighboring regions. When the pixel maintains a similar value for long enough, it is considered as a background pixel, and when its value experiences a large change, it is assumed caused by a foreground passing through. If some part of the foreground is with values similar to the background, the pixel-based models often result in foregrounds with broken segments, which are part of the actual foreground but falsely categorized as background. The broken foregrounds are often seen when the contrast between foregrounds and backgrounds reduces to below some value. This thesis exploits the fact that a foreground appears without broken segments when its contrast to the background is sufficiently large, and proposes a component-based method to compensate the broken segments at low contrast cases using the foreground captured at high contrast occasions. Experiments on videos of various scenes show that the proposed method can improve the segmentation and detection of foregrounds, especially in low contrast scenes.