Information Technology Reference
In-Depth Information
Table 1. Confusion matrix of the recognition results of the first experiment, where 10
positive and 11 negative samples were used for training the SVM. The remaining data
(23 positive and 28 negative) were used for evaluation.
HDG HFD HOG3D
Positive Negative Positive Negative Positive Negative
Positive TP=18 FN=5
TP=19 FN=4
TP=23 FN=0
Negative FP=0
TN=28
FP=3
TN=25
FP=1
TN=27
Table 2. Confusion matrix of the recognition results of the second experiment. In
this test 24 positive and 36 negative samples were used for training the SVM. The
remaining data (48 positive and 94 negative) were used for evaluation.
HDG HFD HOG3D
Positive Negative Positive Negative Positive Negative
Positive TP=46 FN=2
TP=48 FN=0
TP=48 FN=0
Negative FP=1
TN=93
FP=3
TN=91
FP=2
TN=92
Table 3. Computational costs of the different steps in the recognition procedure
HDG extraction
21.18 msec
HFD extraction
48.81 msec
HOG3D extraction
34.30 msec
SVM-based recognition
0.79 msec
the first test, where the HDG -based recognition resulted in a 95.83% TPR (true
positive rate), and a 1.06% FPR (false positive rate), while using the HFD -based
detector a TPR=100% and a FPR=3.19% were achieved. Please note that only
the best results are presented, which were obtained by using the arrangement of
Fig. 1b with 6 and 8 bins for the HFD and the arrangement of Fig. 1a with 6
and 8 bins for the HDG . In the second test by using the increased training set
we obtained FP=1 and FN=0 for the HFD method, FP=1 and FN=2 for the
HDG , FP=3 and FN=0 for the HOG3D .
Finally, we also measured the duration of each step in the recognition proce-
dure. The computation results are summarized in Table 3.
6Con lu on
In this paper we presented a novel approach for recognizing human action. We
used two different spatio-temporal motion-based descriptors and different quanti-
zation arrangements to characterize the event. Instead of representing the action
as a set of features extracted at interest point locations, in our approach one sin-
gle feature describes the whole action. To test our method we used a publicly
 
Search WWH ::




Custom Search