A Motion Activity Descriptor and Its Extraction in Compressed Domain

Xinding Sun*, Ajay Divakaran**, B. S. Manjunath*

* Department of Electrical and Computer
Engineering, University of California,
Santa Barbara, CA 93106
{xdsun, manj} [at] ece.ucsb.edu

** Mitsubishi Electric Research Laboratories,
571 Central Avenue, #115
Murray Hill, NJ 07974
ajayd [at] merl.com

Abstract

A novel motion activity descriptor and its extraction from a compressed MPEG (MPEG-1/2) video stream are presented. The descriptor consists of two parts, a temporal descriptor and a spatial descriptor. To get the temporal descriptor, the "motion intensity" is first computed based on P frame macroblock information. Then the motion intensity histogram is generated for a given video unit as the temporal descriptor. To get the spatial descriptor, the average magnitude of the motion vector in a P frame is used to threshold the macro-blocks into "zero" and "non-zero" types. The average magnitude of the motion vectors and three types of runs of zeros in the frame are then taken as the spatial descriptor. Experimental results show that the proposed descriptor is fast, and that the combination of the temporal and spatial attributes is effective. Key elements of the intensity parameter, spatial parameters and the temporal histogram of the descriptor have been adopted by the draft MPEG-7 standard 10.
[PDF] [BibTex]
X. Sun, A. Divakaran, B. S. Manjunath,
IEEE Pacific-Rim Conference on Multimedia (PCM '01), pp. 450-457, Beijing, China, Oct. 2001.
Node ID: 326 , DB ID: 123 , VRLID: 91 , Lab: VRL , Target: Proceedings