Page 126 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience
P. 126

Section 4.5.  Frequency-Domain  Methods                       103


            where  F t  and  F t−@t  are  the  FTs  of  the  current  and  reference  frames,  respec-
            tively.  In  Ref.  88,  Haskell  noticed  this  relationship  but  did  not  propose  an
            algorithm  to recover the displacement  from the phase  shift.
               If  we  de ne  @ (w x ;w y  )  as  the  phase  di erence  between  the  FT  of  the
            current frame and that of  the reference  frame, then


                           e j@ (w x ;w y  ) = e j[  t  (w x ;w y  )−  t−@t  (w x ;w y  )]
                                    = e j  t  (w x ;w y  )  · e −j  t−@t  (w x ;w y  )
                                       F t  (w x ;w y  )   F  ∗  (w x ;w y )
                                    =           ·   t−@t     ;          (4.24)
                                                   ∗
                                      |F t  (w x ;w y  )| |F t−@t  (w x ;w y )|
            where    t  and    t−@t  are  the  phase  components  of  F t  and  F t−@t  ,  respectively,
            and the superscript ∗ indicates the complex conjugate. If we de ne c t; t−@t  (x; y)
            as the inverse FT of  e j@ (w x ;w y  ) , then

                                 −1  j@ (w x ;w y  )
                   c t; t−@t  (x; y)=  F  {e  }
                                 −1  j  t  (w x ;w y  )   −j  t−@t  (w x ;w y  )
                             = F   {e       · e         }
                                 −1  j  t  (w x ;w y  )  −1  −j  t−@t  (w x ;w y  )
                             = F   {e       }⊗ F   {e          };       (4.25)
            where ⊗ is the 2-D convolution operation. In other words, c t; t−@t  (x; y)isthe
            cross-correlation of the inverse FTs of the phase components of F t  and F t−@t  .
            For  this  reason,  c t; t−@t  (x; y)  is  known  as  the  phase  correlation  function.  The
            importance of this function becomes apparent if it is rewritten in terms of the
            phase di erence  in Equation  (4.23):
                                 −1  j@ (w x ;w y  )
                   c t; t−@t  (x; y)=  F  {e  }
                                 −1  j(−w x d x  −w y d y  )
                             = F   {e          }
                             =  (x − d x ;y  − d y  ):                  (4.26)
            Thus,  the  phase  correlation  surface  has  a  distinctive  impulse  at  (d x ;d y  ).  This
            observation  is  the  basic  idea  behind  the  phase  correlation  motion  estimation
            method.  In  this  method,  Equation  (4.24)  is  used  to  calculate  e j@ (w x ;w y  ) ,  the
            inverse  FT  is  then  applied  to  obtain  c t; t−@t  (x; y),  and  the  location  of  the
            impulse in this function  is  detected to estimate  (d x ;d y ).
               In  practice,  the  impulse  in  the  phase  correlation  function  degenerates  into
            one or more peaks. This is due to many factors, like the use, in digital images,
            of the discrete Fourier transform (DFT) instead of the FT, the presence of more
            than  one  moving  object  within  the  considered  area  A,  and  the  presence  of
   121   122   123   124   125   126   127   128   129   130   131