Browse Prior Art Database

Activity Detection

IP.com Disclosure Number: IPCOM000122611D
Original Publication Date: 1991-Dec-01
Included in the Prior Art Database: 2005-Apr-04
Document File: 3 page(s) / 133K

Publishing Venue

IBM

Related People

Lancon, E: AUTHOR [+2]

Abstract

In a time-prediction video coding loop, Activity Detection (AD) can be used to reduce the temporal redundancy of a video sequence and, therefore, the bit-rate of its coded version. AD is mostly used in combination with Motion Compensation (MC) which compensates for the motions within a sequence, either due to an object in motion or a displacement of the video camera. AD and MC used together lead to high compression gains. However, in case of significant motion of an object or of the video camera, the number of "active blocks", which the AD algorithm detects, becomes very large and lead to increase the computational complexity of the encoder and the output bit-rate.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 49% of the total text.

Activity Detection

      In a time-prediction video coding loop, Activity
Detection (AD) can be used to reduce the temporal redundancy of a
video sequence and, therefore, the bit-rate of its coded version. AD
is mostly used in combination with Motion Compensation (MC) which
compensates for the motions within a sequence, either due to an
object in motion or a displacement of the video camera. AD and MC
used together lead to high compression gains. However, in case of
significant motion of an object or of the video camera, the number of
"active blocks", which the AD algorithm detects, becomes very large
and lead to increase the computational complexity of the encoder and
the output bit-rate.

      In this article we propose to apply a Second Activity Detection
over the time-prediction error signal derived from the original video
sequence to further reduce the temporal redundancy and, therefore,
the computational complexity and the bit-rate of the coded sequence.

      AD consists in dividing each frame of a sequence into small
blocks (generally, 4x4, 8x8 pixels) and to subtract from each block
of each frame, the same block (at the same location) in the previous
frame.

      A psychovisual criterion tells us whether the difference is
above a given threshold or not. If yes, the "active" blocks are
encoded. If not, change from one block to the next in time is not
detectable by the human eye and is therefore not encoded. To
reconstruct the video sequence, the decoder replaces the non-active
blocks by the corresponding blocks of the previous frame, without the
human eye noticing it.  For a "head and shoulder" sequence (fixed
video camera shooting a speaker who does not move too much), about
90% of the blocks (example taken from a sequence of 256x256 images
divided into 1024 blocks of 4x4 pixels) are found non-active, leading
to high compression gains. In addition, Motion Compensation can be
performed on the "active" blocks. Each active block may result from
the motion of an object in the sequence or of the video camera. From
each block is subtracted a block in the previous frame displaced by a
motion vector that has been estimated, and which corresponds to the
motion of an object in that block or of the video camera. The better
the two blocks match, the lower their difference, the better the
coding gain. Those blocks are the "motion compensation error blocks"
that result from the reduction of the temporal redundancy of a video
sequence using AD and MC. They are then encoded by reducing the
spatial redundancy in each of them, to further compress the video
signal. The coded information finally consists in the spatially coded
motion compensated error blocks and the motion vectors.

      In a few cases, after motion compensation, the two blocks
perfectly match, the error block is zero and does not need to be
spatially encoded (although it was found active by the AD process).
Only the information about the motion vectors is thus...