Page 141 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

118 Chapter 4. Basic Motion Estimation Techniques

DFA This is an implementation of the di erential method of Ca orio and
Rocca as given by Equation (4.12). In this case, the moving area, A,
was set to a block of 16 × 16 pels.
PRA This is an implementation of the pel-recursive algorithm of Netravali
and Robbins as given by Equation (4.20). In this case, the motion vector
ˆ i
of the previous pel in the line was taken as the initial motion estimate, d ,
of the current pel, the update step size was set to =1=1024, the update
term was calculated and averaged over an area of 3 × 3 pels centered
around the current pel, and ve iterations were performed per pel.
PCA This is an implementation of the phase correlation method as given
by Equations (4.24) and (4.25). In this case, a window of 32 × 32 pels
centered around the current 16 × 16 block was used to generate the phase
correlation surface. The three most dominant peaks in this surface were
detected and the corresponding motion displacements were unwrapped
using Equation (4.27). The three candidate displacements were then tested
using the SAD between the current block and the candidate displaced
block in the reference frame. The candidate displacement with the lowest
SAD was chosen as the motion vector of the current block.
BMA This is an implementation of a full-search block-matching algorithm.
In this case, the block size was 16 × 16 pels and the matching criterion
was the SAD.
In each case, the maximum allowed motion displacement was set to
± 15 pels in each direction and the motion vectors were allowed to point
outside the reference frame (i.e., unrestricted motion vectors). To provide a
fair comparison and to ease motion vector coding, all displacements were
estimated with half-pel accuracy. In DFA and PRA this was achieved by
rounding the subpel accurate motion estimates to the nearest half-pel accu-
rate motion vectors. In PCA and BMA this was achieved using a re nement
stage that examined the eight nearest half-pel estimates centered around the
full-pel motion estimate. Bilinear interpolation was used to obtain intensity
values at subpel locations of the reference frame. To mask the e ect of the
temporal propagation of prediction errors, motion was estimated and com-
pensated using original reference frames. For comparison purposes, motion
vectors were coded using the median predictor and the VLC table of the
H.263 standard. The DFD signal was also transform encoded according to
the H.263 standard and a quantization parameter of QP =10. All quoted re-
sults refer to the luma components of sequences. No chroma encoding was
performed.
Care should be taken when interpreting the results of this study. Di erent
simulation parameters will lead to di erent results. For example, at the expense