A fast and accurate video object detection and segmentation method in the compressed domain

Abstract

In this paper, we present a fast and accurate method for object detection and segmentation in the compressed domain. First the motion vectors are emendated by our spatial confidence and correction rule so that these can represent real motions of objects. Then the initial location and a coarse segmentation from the motion vectors are obtained by applying the EM algorithm. For there are DCT coefficients only in I frames, the coarse masks are mapped into I frames through the motion parameters. These blocks in the masks can be decompressed to obtain details of a specific object in the pixel domain. The actual edges of the objects can be extracted by applying Canny Edge Detection only in the segmented regions. By using the proposed algorithm, the amount of data needed to be processed is kept the necessarily minimal, saving the computational time as well as gaining the pixel-wise edges of the segmented objects.