Segmenting Moving Objects in MPEG Videos in the Presence of Camera Motion

Abstract

The distinction of translational and rotational camera motion and the recognition of moving objects is an important topic for scientific film studies. In this paper, we present an approach to distinguish between camera and object motion in MPEG videos and provide a pixel-accurate segmentation of moving objects. Compressed domain features are used as far as possible in order to reduce computation time. First, camera motion parameters are estimated and translational movements are distinguished from rotational movements based on a three-dimensional (3D) camera model. Then, motion vectors which do not fit to the camera motion estimate are assigned to object clusters. The moving object information is utilized to refine the camera motion estimate, and a novel compressed domain tracking algorithm is applied to verify the temporal consistency of detected objects. In contrast to previous approaches, the tracking of both moving objects and background allows to perform their separation iteratively only once per shot. The object boundary is estimated with pixel accuracy via active contour models. Experimental results demonstrate the feasibility of the proposed algorithm.