3。2。3。 Gradient descent optimization。 Computing an auto-correlation SSD surface can be a time-consum- ing operation。 Since the system will be performing this operation for each candidate feature window, we should try to limit the number of candidates that are considered for each object of interest。 Although an exhaustive search of each possible feature window placement within an object’s bounding box  would find the highest confidence value, it may be desirable to sacrifice this guarantee in favor of finding a confidence measure that may not always be ideal, but can   generally   be   found   faster。   This   is  done by

performing  a  gradient  descent  search26 through the

space of candidate feature window placements。 Beginning at an arbitrary window placement, the search consists of a series of locally optimal greedy decisions。 Each decision computes the confidence measure of neighboring window placements and selects the best (the maximum)。 At the selected location, the same decision process is repeated until no neighboring confidence measure is better than the current one。

4。 FIGURE  SEGMENT  CORRESPONDENCES The detection module provides no correspondence among the figure segments representing an object at different points in time。 The most straightforward approach  to corresponding segments is to establish  a

correspondence  if  and  only  if  a segment’s  bounding    correlation   technique   can  be  used   to correspond box intersects the bounding box of a segment in the figure segments for the visual tracking of objects。 previous iteration。 However, the  adjacency  method  Instead  of  taking  the  window  li„ (,j  to  be  a  figure imposes  a limit  on  the total amount  of time  that  an  segment’s  bounding   box,   it   can   be   a   small, iteration  of the system can take。 If a moving object is   recognizable   portion  of  that  object  (i。e。  a    feature able  to  move  to  a  location  where  its  bounding  box   window)。   Again,   the   tracking   module  performs does  not  intersect  the  bounding  box   from   the   several   iterations   of   SSD-produced   displacement previous  iteration, then the tracking system   fails。 measurements,   after   which   the   detection  module

locates   new   figure   segments。   Object   identity   is

4。 l 。  Cross-correlation techniques

We have experimented with an alternative method, where figure segments are visually tracked by performing several iterations of object cross-correla- tion。 During each inpidual iteration of object cross- correlation, figure segment displacements are com- puted through the use of optical flow。 Unlike the derivation of auto-correlation in Eq。 (9), the intensity values for I (x, y, t) are obtained  from  an earlier image frame that was stored in memory。 The SSD error measure of Eg。 (8) is computed for displace- ments from the neighborhood N p ,(t) to form a cross- correlation surface:

( J 0)

where the window lF„ , used to calculate the SSD measure is the feature segment’s bounding box。 The displacement that produces the minimum s, value is then taken to be the displacement of the figure segment。 After performing  a predetermined number of iterations, the detection module is called  to produce a new set of figure segments。 If a detected segment is found to intersect one that has been tracked, then it is assumed to correspond to the same object of interest。