As a sequel to the JPEG standards committee, the Moving Picture Experts Group (MPEG) was set up in the mid 1980s to agree standards for video sequence compression.

Their first standard was MPEG-1, designed for CD-ROM applications at 1.5 Mb/s, and their more recent standard, MPEG-2, is aimed at broadcast quality TV signals at 4 to 10 Mb/s and is also suitable for high-definition TV (HDTV) at 20 Mb/s. We shall not go into the detailed differences between these standards, but simply describe some of their important features. MPEG-2 is used for digital TV and DVD in the UK and throughout the world.

MPEG coders all use the MCPC structure of this previous figure, and employ the 8×8 8 8 DCT as the basic transform process. So in many respects they are similar to H.261 coders, except that they operate with higher resolution frames and higher bit rates.

The main difference from H.261 is the concept of a Group of Pictures (GOP) Layer in the coding hierarchy, shown in Figure 1 . However we describe the other layers first:

• The Sequence Layer contains a complete image sequence, possibly hundreds or thousands of frames.
• The Picture Layer contains the code for a single frame, which may either be coded in absolute form or coded as the difference from a predicted frame.
• The Slice Layer contains one row of macroblocks ( 16×16 16 16 pels) from a frame. (48 macroblocks give a row 768 pels wide.)
• The Macroblock Layer contains a single macroblock -- usually 4 blocks of luminance, 2 blocks of chrominance and a motion vector.
• The Block Layer contains the DCT coefficients for a single 8×8 8 8 block of pels, coded almost as in JPEG using zig-zag scanning and run-amplitude Huffman codes.

The GOP Layer contains a small number of frames (typically 12) coded so that they can be decoded completely as a unit, without reference to frames outside of the group. There are three types of frame:

• Intra coded frames (I) -- which are coded as single frames as in JPEG, without reference to any other frames.
• Predictive coded frames (P) -- which are coded as the difference from a motion compensated prediction frame, generated from an earlier I or P frame in the GOP.
• Bi-directional coded frames (B) -- which are coded as the difference from a bi-directionally interpolated frame, generated from earlier and later I or P frames in the sequence (with motion compensation).

The main purpose of the GOP is to allow editing and splicing of video material from different sources and to allow rapid forward or reverse searching through sequences. A GOP usually represents about half a second of the image sequence.

Figure 2 shows a typical GOP and how the coded frames depend on each other. The first frame of the GOP is always an I frame, which may be decoded without needing data from any other frame. At regular intervals through the GOP, there are P frames, which are coded relative to a prediction from the I frame or previous P frame in the GOP. Between each pair of I / P frames are one or more B frames.

The I frame in each GOP requires the most bits per frame and provides the initial reference for all other frames in the GOP. Each P frame typically requires about one third of the bits of an I frame, and there may be 3 of these per GOP. Each B frame requires about half the bits of a P frame and there may be 8 of these per GOP. Hence the coded bits are split about evenly between the three frame types.

B frames require fewer bits than P frames mainly because bi-directional prediction allows uncovered background areas to be predicted from a subsequent frame. The motion-compensated prediction in a B frame may be forward, backward, or a combination of the two (selected in the macroblock layer). Since no other frames are predicted from them, B frames may be coarsely quantised in areas of high motion and comprise mainly motion prediction information elsewhere.

In order to keep all frames in the coded bit stream causal, B frames are always transmitted after the I/P frames to which they refer, as shown at the bottom of Figure 2 .

One of the main ways that the H.263 (enhanced H.261) standard is able to code at very low bit rates is the incorporation of the B frame concept.

Considerable research work at present is being directed towards more sophisticated motion models, which are based more on the outlines of objects rather than on simple blocks. These will form the basis of extensions to the new low bit-rate video standard, MPEG-4 (MPEG-3 is an audio coding standard).