Survey of Dirac: A Wavelet Based Video Codec for Multiparty Video Conferencing and Broadcasting - Intelligent Video Event Analysis and Understanding - page 243

reduce the computational intensiveness by improving the ME algorithm. In [3],
Tun et al. proposed a fast and efficient motion estimation algorithm which is based
on semi hierarchical approach. It combines hierarchical and semi-hierarchical
approaches for different types of inter frames. In [14], Khan et al. proposed a
motion estimation algorithm for real-time H.264 video encoder, which is based on
three dimensional recursive search (3DRS) algorithm. We have worked to
improve the motion estimation module by replacing the current motion estimator
with 3DRS [13]. In [4], we have implemented only the IPPP… structure. We have
extended our approach to incorporate B frames. Now the structure will be
IBBPBBP…. Like original 3DRS algorithm [13], the five candidate blocks
involved in the initial prediction of motion vectors are as shown in Figure 34.

A

B

MB in current frame

C X

MB in previous frame

Current MB

E

Fig. 34 Neighboring macro blocks used as candidates

The proposed Scheme is as shown in the Figure 35. The forward direction arrow
in Figure 35 indicates that the temporal prediction MVs will be positive and a
backward direction arrow indicates that they should be negative.

Fig. 35 Proposed Motion Estimation Scheme

It should also be noted that the best MVs from the P frames cannot be used
directly as a temporal prediction since the prediction structure is different for both
P and B as shown in Figure 35. The temporal prediction MVs for B frames
requires scaling up or down since the temporal distances of both references are
different for B frames. For example, the temporal distance of B1 to its 2nd
reference is 2 and the temporal distance of P1 to its 1st reference is 3, so in order
to use P1 MVs as a temporal candidate vectors, it should be multiplied by -2/3. -
ve sign is used as the vectors are used for backward frame. Similarly, for B2 the
MVs of P1 are multiplied by -1/3. It is important to note that the temporal
predicted MVs are available only for the level 0 motion estimation and are uptil 1
pixel accuracy.