Digital Signal Processing Reference
In-Depth Information
2.
Add the spatial and temporal prior to each pixel, and calculate the local energy of
each pixel by combining the prior and the likelihood;
3.
Compute the global energy, and if the increment is small than a threshold (set 10),
stop to get the final labeling field, otherwise, go to 2 to continue.
5
Experiment Results
Two typical dynamic backgrounds, waving trees and water rippling from the new
dataset [19], are used to evaluate the effectiveness of the proposed method in com-
parison with GMM and T2-FGMM. The parameters of the algorithms are shown in
Table. 1 (the parameters for GMM and T2-FGMM are from the works [4], [8, 9]).
Table 1.
Parameters of GMM, T2-FGMM and our method
α
ξ
ς
β
c
K
β
'
GMM
3
0.001
0.25
--
--
--
--
T2-FGMM
3
0.001
0.25
2
2.5
--
--
Our method
3
0.001
0.25
2
2.5
0.9
2.8
To evaluate the methods in pixel level, false positive (FP), false negative (FN), true
positive (TP), true negative (TN) are used. P denotes the background, and N denotes
the foreground. T presents the judgment of correct classification, and F presents the
judgment of wrong classification. Statistical metrics give an overall performance in a
much fairer way in dynamic backgrounds. False positive rate (FPR), Percentage of
Bad Classification (PBC), F-measure are given as follows:
FP
FN+FP
FPR
=
, PBC
=
100
×
FP+TN
FN+FP+TN+TP
(6)
TP
TP
recall
×
precision
recall
=
, precision
=
, F
=
2
×
.
TP+FN
TP+FP
recall
+
precision
5.1
Waving Trees
We use the “overpass” sequence to test the waving trees situation. Fig. 1 shows the
input images, ground truth and the output of GMM, T2-FGMM and the proposed
approach. The evaluation of these three approaches in pixel level is shown Table. 2
and Fig. 3. The results show that our method can restrain the dynamic background
better than GMM and T2-FGMM. At the same time, the overall performance is better
than GMM and T2-FGMM as well.
