TW200743393A - Method of real-time hierarchical background reconstruction and foreground detection - Google Patents

Abstract

Disclosed is a method of real-time hierarchical background reconstruction and foreground detection which is capable of reconstructing background and detecting foreground during video communication and is characterized in integrating block-based coarse-level background modeling and pixel-based fine-level background modeling. The method comprises steps of the followings: dividing an image into non-overlap blocks and establishing coarse-level background modeling for each block; performing rapid but coarse detection for foreground objects to generate a foreground detection result; and feeding the foreground detection result to the fine-level background modeling for further analysis. A new-developed block descriptor in the coarse-level background reconstruction part is used to describe characteristics of information contained in the blocks and the descriptor is called a difference mean vector (DMV). In addition, a Gaussian mixture method is used to construct coarse-level background modeling for DMV characteristics. The fine-level parts can adapt conventional pixel-based background modeling. A feed-forward framework is also disclosed to integrate coarse-level background modeling and fine-level background modeling by feeding information identified in coarse-level into the fine-level for fining the coarse-level foreground detection result and updating the fine-level background modeling.