3.2. Feature window selection

Before addressing the problem of automatically selecting feature windows, it is necessary to under- stand what qualities we desire the feature window to possess. For our serveing task, we would like feature windows to have the follwing  two  important qualities:

• The feature window should contain enough  pixels that correspond to the object of interest so that the window may be effectively used for tracking. Ideally, each of the feature window’s pixels correspond to the object,  causing  the  window to be invariant with respect to the object’s surround- ings.

• The pixels within the feature window must possess enough  contrast  for the window  to  be tracked.

(a-b) The user initially selects a sponta- neous domain representing the beginning of the conveyor belt.

(c-e) As the target appears, its continuous domain initially intersects the spontaneous domain. The two domains are combined into one.

(f-g) As the object moves away, the two   domains eventually separate.

Fig. 4. Segmentation domains.

By computing a bounding box around an object   of critical  because  visual  tracking  algorithms  may fail interest,  we  have  provided  a  solution  for  the  first of due to  repeated  patterns  or  areas  of  uniform intensity these two     concerns.     The   problem of    locating in  the  intensity  arrangement.  Both  situations  cause  a trackable  features  for  an  object  has  been   reduced tracking algorithm to reach ambiguous matches  from from   a  search   throughout   the   entire   image   to a within   its   search   neighborhood,   resulting   in   an search within  the object’s bounding  box. output  of  spurious  displacement  measures.  In order The   qualification   of   feature   window   intensity to  avoid   this  problem,   our  system  considers  and

arrangement   with   a   good   confidence   measure is evaluates  many  candidate  placements  of  a   feature

window.  Our  system  then  selects  feature   window 3.2.2. Confidence measures. Feature selections  made placements   that   optimize   the   performance   of   a from  within  a  bounding  box  typically  lie   entirely tracking algorithm. within  the object  or contain  at least  a portion  of the object’s  pixels.   Confidence  measures   are  used  to

determine   which   of   these   are   most   suited    for

3.2. 1. Use of optical flow