def getnotassociatedindex(len_sub, len_tr, del_tr, del_sub):
non_tr = []
non_sub = []
non_sub = np.array(range(len_sub))
non_tr = np.array(range(len_tr))
non_sub = np.delete(non_sub, del_sub)
non_tr = np.delete(non_tr, del_tr)
non_sub = non_sub.tolist()
non_tr = non_tr.tolist()
return non_sub, non_tr
def getnotassociatedindex(len_sub, len_tr, del_tr, del_sub):
non_tr = []
non_sub = []
non_sub = np.array(range(len_sub))
non_tr = np.array(range(len_tr))
non_sub = np.delete(non_sub, del_sub)
non_tr = np.delete(non_tr, del_tr)
non_sub = non_sub.tolist()
non_tr = non_tr.tolist()
return non_sub, non_tr
def trackmerge(tr, new_tr_copy, non_tr, loss, threshold, res):
threshold_ = threshold + (2 * threshold / 3) # margin
new_tr = new_tr_copy
for ii in range(len(non_tr)):
a = loss[non_tr[ii], :]
b = a[a < threshold_]
if len(b) > 0:
if len(b) > 1:
b = [b.min()]
# Get merging track overlapped subject's index
idx_b = np.argwhere(a == b)
# Get parent track's index
idx_new_tr = np.argwhere(res[:, 1] == idx_b[0, 0])
# Merge tracks
try: # Maybe wrong hungarian as we expected
# Only associate locked tracks
if tr[ii].state is 2 and new_tr[idx_new_tr[0, 0]].state is 2:
new_tr[idx_new_tr[0, 0]].associatetrack(tr[ii])
tr = np.delete(tr, ii)
tr = tr.tolist()
except:
pass
return new_tr, tr
def trackmerge(tr, new_tr_copy, non_tr, loss, threshold, res):
threshold_ = threshold + (2 * threshold / 3) # margin
new_tr = new_tr_copy
for ii in range(len(non_tr)):
a = loss[non_tr[ii], :]
b = a[a < threshold_]
if len(b) > 0:
if len(b) > 1:
b = [b.min()]
# Get merging track overlapped subject's index
idx_b = np.argwhere(a == b)
# Get parent track's index
idx_new_tr = np.argwhere(res[:, 1] == idx_b[0, 0])
# Merge tracks
try: # Maybe wrong hungarian as we expected
# Only associate locked tracks
if tr[ii].state is 2 and new_tr[idx_new_tr[0, 0]].state is 2:
new_tr[idx_new_tr[0, 0]].associatetrack(tr[ii])
tr = np.delete(tr, ii)
tr = tr.tolist()
except:
pass
return new_tr, tr
def hungarianassociation(loss, threshold):
print loss
# SKLEARN association method
res = _hungarian(loss)
del_index = []
for ii in range(len(res)):
y, x = res[ii]
"""
if(loss[y, x] > threshold):
del_index.append(ii)
"""
new_res = np.delete(res, del_index, 0)
return new_res
def hungarianassociation(loss, distance, threshold):
loss = postproc(loss, threshold)
debug_flag = 0
if debug_flag: print loss
# SKLEARN association method
res = _hungarian(loss)
del_index = []
for ii in range(len(res)):
y, x = res[ii]
if(loss [y, x] > threshold):
del_index.append(ii)
new_res = np.delete(res, del_index, 0)
#print new_res
return new_res
def hungarianassociation(loss, distance, threshold):
loss = postproc(loss, threshold)
debug_flag = 0
if debug_flag: print loss
# SKLEARN association method
res = _hungarian(loss)
del_index = []
for ii in range(len(res)):
y, x = res[ii]
if (loss[y, x] > threshold):
del_index.append(ii)
new_res = np.delete(res, del_index, 0)
#print new_res
return new_res
def tracksplit(new_tr, sub, threshold):
# usage of appareance model might be a good option
# for distance calculation in this section
threshold_ = threshold + (2 * threshold / 3) # margin
del_idx = []
for ii in range(len(sub)):
for tr in new_tr:
if tr.group and tr.calculatesubjectdistance(sub[ii], threshold_):
n_tr = tr.deassociatetrack()
n_tr.updatetrack(sub[ii])
new_tr.append(n_tr)
del_idx.append(ii)
break
if del_idx:
sub = np.delete(sub, del_idx)
sub = sub.tolist()
return new_tr, sub
def tracksplit(new_tr, sub, threshold):
# usage of appareance model might be a good option
# for distance calculation in this section
threshold_ = threshold + (2 * threshold / 3) # margin
del_idx = []
for ii in range(len(sub)):
for tr in new_tr:
if tr.group and tr.calculatesubjectdistance(sub[ii], threshold_):
n_tr = tr.deassociatetrack()
score = lossfunction(tr, sub[ii])
n_tr.updatetrack(sub[ii], score[0])
new_tr.append(n_tr)
del_idx.append(ii)
break
if del_idx:
sub = np.delete(sub, del_idx)
sub = sub.tolist()
return new_tr, sub
def trackupdate(tr, sub, res, loss, threshold):
new_track = []
del_index_sub = []
del_index_tr = []
init_tr = tr
init_sub = sub
threshold_distance = 100
aux_res = res
# Update successful associations
for ii in range(len(res)):
y, x = res[ii]
aux_res = np.delete(aux_res, 0, axis=0)
new_tr = assignsubjecttoexistingtrack(tr[y], sub[x])
new_track.append(new_tr)
del_index_sub.append(x)
del_index_tr.append(y)
sub = np.delete(sub, del_index_sub)
tr = np.delete(tr, del_index_tr)
sub = sub.tolist()
tr = tr.tolist()
# Update missed associations --> where merge should act
non_index_sub, non_index_tr = getnotassociatedindex(
len(init_sub), len(init_tr), del_index_tr, del_index_sub)
new_track, tr = trackmerge(
tr, new_track, non_index_tr, loss, threshold_distance, res)
del_index = []
for ii in range(len(tr)):
tr[ii].updatetrack()
# In case track got lost
if tr[ii].state == 4:
del_index.append(ii)
if del_index:
tr = np.delete(tr, del_index)
tr = tr.tolist()
for n in new_track:
tr.append(n)
# Update new subjects --> where split should act
tr, sub = tracksplit(new_track, sub, threshold_distance)
"""
TODO
----
- Only assign new tracks in outer positions of the image
"""
for s in sub:
t = assignsubjecttonewtrack(s)
tr.append(t)
return tr
def trackupdate(tr, sub, res, loss, threshold):
new_track = []
del_index_sub = []
del_index_tr = []
init_tr = tr
init_sub = sub
threshold_distance = 100
aux_res = res
# Update successful associations
for ii in range(len(res)):
y, x = res[ii]
aux_res = np.delete(aux_res, 0, axis=0)
new_tr = assignsubjecttoexistingtrack(tr[y], sub[x])
new_track.append(new_tr)
del_index_sub.append(x)
del_index_tr.append(y)
sub = np.delete(sub, del_index_sub)
tr = np.delete(tr, del_index_tr)
sub = sub.tolist()
tr = tr.tolist()
# Update missed associations --> where merge should act
non_index_sub, non_index_tr = getnotassociatedindex(
len(init_sub), len(init_tr), del_index_tr, del_index_sub)
new_track, tr = trackmerge(tr, new_track, non_index_tr, loss,
threshold_distance, res)
del_index = []
for ii in range(len(tr)):
tr[ii].updatetrack()
# In case track got lost
if tr[ii].state == 4:
del_index.append(ii)
if del_index:
tr = np.delete(tr, del_index)
tr = tr.tolist()
for n in new_track:
tr.append(n)
# Update new subjects --> where split should act
tr, sub = tracksplit(new_track, sub, threshold_distance)
"""
TODO
----
- Only assign new tracks in outer positions of the image
"""
for s in sub:
t = assignsubjecttonewtrack(s)
tr.append(t)
return tr
