Page 523 - Fundamentals of Radar Signal Processing
P. 523

about 65 percent to avoid increasing the CFAR loss. Furthermore, for N > 8 the

               log  CFAR  loss  in  decibels  is  about  65  percent  more  than  the  loss  for  a  CA
               CFAR with the same value of N.


               6.5.6   Order Statistic CFAR
               An  alternative  to  cell-averaging  CFAR  is  the  class  of rank-based  or order
               statistic  CFARs  (OS  CFAR).  Proposed  primarily  for  combating  masking
               degradations, OS CFAR retains the one-dimensional or two-dimensional sliding

               window structure of CA CFAR, including guard cells if desired, but does away
               entirely with averaging of the reference window contents to explicitly estimate
               the interference level. Instead, OS CFAR rank orders the reference window data
               samples {x , x ,…, x } to form a new sequence in ascending numerical order,
                                 2
                            1
                                        n
               denoted by {x , x , …, x }. The kth element of the ordered list is called the
                                (1)
                                     (2)
                                               (N)
               kth order statistic. For example, the first order statistic is the minimum, the Nth
               order statistic is the maximum, and the (N/2)th order statistic is the median of
               the  data  {x ,  x ,…, x }.  In  OS  CFAR,  the kth  order  statistic  is  selected  as
                                         N
                             1
                                 2
               representative of the interference level and a threshold is set as a multiple of
               this value:


                                                                                                     (6.158)


               The interference is thus estimated from only one actual data sample, instead of

               an average of all of the data samples. Nonetheless, the threshold in fact depends
               on all of the data since all of the samples are required to determine which will
               be the kth largest.
                     It will be shown that this algorithm is in fact CFAR (i.e., does not depend

               on the interference power           ), and the threshold multiplier required to achieve
               a  specified   will be determined. The analysis follows (Levanon 1988). To

               simplify the notation, consider the square-law detected output x  normalized to
                                                                                            i
               its mean,            ; this will have an exponential PDF with unit mean. The rank-
               ordered  set  of  reference  samples  {y }  are  denoted  by  {y }.  For  a  given
                                                              i
                                                                                         (i)
               threshold T, the probability of false alarm will be




                                                                                                     (6.159)

               The average P  will be computed as
                                FA




                                                                                                     (6.160)
   518   519   520   521   522   523   524   525   526   527   528