Recently, the content-based video coding schemes are needed for many applications. Region-of-interest extraction is important to these coding algorithms. Especially, region-of-interest extraction is essential for the automated video surveillance system. Video surveillance has two main differences compared with other applications: the pre-acquirable background information due to the restricted spatial view of camera and the frequent variations of illumination.
This paper presents the algorithm adequate to this video surveillance. The proposed method is composed of three steps: image indexing part, image alignment part and region-of-interest decision part. The background model varies as an input frame caused by a moving camera. The background model for an input frame is indexed efficiently from the pre-defined background sequence. As the camera views of the indexed background frame and an input are different slightly, two frames are aligned by the proposed block matching algorithm. Also, DCT coefficients that have edge characteristics are used for illumination invariant block-level detection.
Experimental results show that the proposed method can extract region-of-interest by the indexed background frame with the illumination changeable environment and moving camera.
