Information Technology Reference
In-Depth Information
by the choice of the threshold and we have chosen a suitable threshold to
have 80 detected STIPs for this comparison. There is also to mention that
Laptev's executable code is compiledinCenvironment,whileourLBP-TOP
implementation is compiled in Matlab environment. Similar performance to
Laptev's is achieved using the Extended LBP-TOP descriptor which is al-
most 3 times computationally faster than the Extended Gradient LBP-TOP
descriptor.
Tabl e 5 Accuracy and computational time for different LBP-TOP methods and
HOG-HOF, k=1000 visual words
Descriptor
length
Computational
time (s)
Accuracy
(SVM)
Method
LBP-TOP
8
,
8
,
8
,
2
,
2
,
2
768
0.0139
86.25 %
Ext Grad LBP-TOP
8
,
8
,
8
,
2
,
2
,
2
2304
0.0992
90.72 %
Ext Grad LBP-TOP
8
,
8
,
8
,
2
,
2
,
2
+
PCA
100
0.1004
91.25 %
HOG-HOF
162
0.2820*
89.88 %
HOG-HOF + PCA
100
0.2894*
89.28 %
5Conclu ion
In this chapter, we have applied LBP-TOP as a descriptor of small video-
patches used in a part-based approach for human action recognition. We
have shown that LBP-TOP descriptor can be suitable for the description of
cuboids extracted from a video sequences and containing information about
human movements and actions.
We have modified the original descriptor introducing the CSLBP-TOP de-
scriptor and we applied the LBP and CS-LBP operator to the original, gradi-
ent and Gabor images. Moreover, we extended LBP-TOP and CSLBP-TOP
considering the action at three different frames in XY plane and at different
views in XT and YT planes.We have also shown that the performance of
descriptor is quite stable when the PCA is applied.
The use of Extended Gradient LBP-TOP permits us to reach the best
results on the KTH human action database by achieving 92.69% classification
accuracy and 92.57% if PCA is applied using 1-NN classifier with
χ
2
distance
and setting the codebook's size equal to 1250. If SVM classifier is chosen, the
classification accuracy is slightly lower, 91.46 % and 91.34% if PCA is applied.
In Figure 16 the confusion matrices are shown for both classifiers. Most of the
confusion happens between the classes running and jogging, as these actions
are very similar to each other, while all actions performed by hands and arms
are quite accurately classified.
Search WWH ::
Custom Search