Spatio-temporal Dynamic Texture Descriptors for Human Motion Recognition - Intelligent Video Event Analysis and Understanding - page 83

Tabl e 3 Accuracy achieved by the SVM classier for the Original and Extended
LBP-TOP

Accuracy
(SVM)

Descriptor
length

Computational
time (s)

Method

LBP-TOP 8 , 8 , 8 , 2 , 2 , 2 86.25 % 768 0.0139

Extended LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
(3 slices on each axis)

88.19 % 2304 0.0314

Extended LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
(3 slices on each axis) +PCA

87.87 % 100 0.0319

Extended LBP-TOP 8 , 8 , 8 , 2 , 2 , 2
(6 slices on each axis) +PCA

88.38 % 100 0.0630

auspicable, since the entire system benefits in computational time and mem-
ory resources: the creation of the codebook (k-means algorithm), the his-
togramming of spatial-temporal words, the training and testing of 1-NN and
SVM are faster. Figure 11 shows the performances of Extended LBP-TOP
descriptor using the original feature vector (2304 feature length) and using
the reduces feature vector with PCA and OLPP (100 feature length). PCA
technique is very useful, since the performances are not decreasing. In our
tests, OLPP dimensionality reduction technique is not giving good results.
As we can see, the performance decreases a lot. Although this technique re-
lies in reflecting the intrinsic geometric structure of the given data space, is
supervised and has been proved to be effective in facial expression analysis
[17], in our human action recognition task it is not useful.

(a) 1-NN classifier (b) SVM classifier

Fig. 11 Dimensionality reduction techniques (PCA and OLPP) applied to
descriptors

Figures 12a and 12b show a comparison for different enhancement of LBP-
TOP, as the codebook's size is increasing. The use of LBP-TOP applied
to the gradient cuboid gives better results compared with the original one.
The information extracted from the gradient calculated along x , y and t