def cc(self,t):
# perform background subtraction
(dfore,bw) = self.bg.sub_bg(t+params.start_frame)
# for each pixel, find the target it most likely belongs to
(y,x) = num.where(bw)
mind = num.zeros(y.shape)
mind[:] = num.inf
closest = num.zeros(y.shape)
for targ in self.tracks[t].itervalues():
S = est.ell2cov(targ.major,targ.minor,targ.angle)
Sinv = num.linalg.inv(S)
xx = x.astype(float) - targ.x
yy = y.astype(float) - targ.y
d = xx**2*Sinv[0,0] + 2*Sinv[0,1]*xx*yy + yy**2*Sinv[1,1]
issmall = d <= mind
mind[issmall] = d[issmall]
closest[issmall] = targ.identity
# set each pixel in L to belong to the closest target
L = num.zeros(bw.shape)
L[bw] = closest+1
#mpl.imshow(L)
#mpl.show()
return (L,dfore)
def splitobservation(bw,dfore,k,init):
(r,c) = num.where(bw)
if DEBUG: print 'number of pixels in component being split: %d'%len(r)
x = num.hstack((c.reshape(c.size,1),r.reshape(r.size,1)))
w = dfore[bw]
if DEBUG: print 'data being clustered: '
if DEBUG: print x
if DEBUG: print 'with weights: '
if DEBUG: print w
# create means and covariance matrices to initialize
mu0 = num.zeros((k,2))
S0 = num.zeros((k,2,2))
priors0 = num.zeros(k)
for i in range(k):
if DEBUG: print 'predicted ellipse %d: '%i + str(init[i])
mu0[i,0] = init[i].x
mu0[i,1] = init[i].y
S0[:,:,i] = est.ell2cov(init[i].major,init[i].minor,init[i].angle)
priors0[i] = init[i].area
(tmpmajor,tmpminor,tmpangle) = est.cov2ell(S0[:,:,i])
priors0 = priors0 / num.sum(priors0)
if DEBUG: print 'initializing with '
if DEBUG: print 'mu0 = '
if DEBUG: print mu0
if DEBUG: print 'S0 = '
if DEBUG:
for i in range(k):
print S0[:,:,i]
if DEBUG: print 'priors0 = '
if DEBUG: print priors0
# are there no data points?
if len(r) == 0:
return None
(mu,S,priors,gamma,negloglik) = kcluster.gmmem(x,mu0,S0,priors0,weights=w,thresh=.1,mincov=.015625)
obs = []
for i in range(k):
(major,minor,angle) = est.cov2ell(S[:,:,i])
obs.append(ell.Ellipse(mu[i,0],mu[i,1],minor,major,angle))
obs[-1].compute_area()
return obs
