Study On Motion Detection In Dynamic Scene

ABSTRACT:As one of the most important subjects in the area of computer vision, detecting moving object in dynamic scene, which means that camera moves, rotates or zooms while monitoring, is of profound value in research and application of visual navigation, security surveillance, intelligent transportation, etc.This paper mainly studies the detection of moving object in dynamic scene. An electronic image stabilization method based on multi-resolution hierarchical motion estimation is proposed to remove camera motion, and then moving object is detected based on blocked adaptive Gaussian Mixture Model (GMM) after image stabilizing. Experiments have been performed for evaluations.For video sequences produced by jittering camera system, a multi-resolution hierarchical motion estimation algorithm for electronic image stabilization is proposed, which applies gray projection algorithm, block matching algorithm and feature matching method on each resolution layer of image respectively. Experimental results demonstrate the advantage of our algorithm over traditional motion estimation algorithm in dealing with high sensitivity to external disturbance and handling large scale displacement or rotation of camera motions.For removing the influence of moving objects in motion estimation computation, a block-based regional estimation algorithm is proposed, which discards the blocks containing moving objects in motion estimation computation and calculates motion vectors only in the areas of blocks containing no moving objects on each resolution layer of images. Experimental results show the efficiency of our approach.For detecting moving objects in video sequences after image stabilization, a block-based adaptive GMM motion detection algorithm is proposed. Firstly the original images are partitioned into blocks, and the blocks are then classified into foreground or background by calculating the Mean Absolute Difference (MAD), and then adaptive GMM based detection of moving objects is implemented only on the foreground blocks. Experimental results show that our approach is capable to deal with the remaining errors of electronic image stabilization in motion detection.