Information Technology Reference
In-Depth Information
Table 3.
Confusion matrix
Bend Jack Pjump Jump Run Side Skip Walk Wave1 Wave2
Bend
9
Jack
9
Pjump
9
Jump
7
2
Run
7
1
1
Side
8
1
Skip
1
1
7
Walk
1
8
Wave1
9
Wave2
1
8
Table 4.
Comparative of our method against others found in the literature
Our method Zhou et al. [4] Ali et al. [9]
Bend
100.0
100.0
100.0
Jack
100.0
95.7
100.0
Pjump
100.0
87.5
55.0
Jump
77.7
92.2
100.0
Run
77.7
86.7
88.8
Side
88.8
86.4
88.8
Skip
77.7
84.0
100.0
Walk
88.8
100.0
100.0
Wave1
100.0
100.0
100.0
Wave2
88.8
89.4
100.0
Reported mean accuracy
90.32
91.4
92.6
results, but our approach uses the simplest feature set and performs in real time.
Performance results are not available in [4] and [9] but in our opinion they are
very computational demanding methods which would be penalized for real time
execution.
5 Conclusions and Future Work
This paper describes a human action recognition system based on a robust vi-
sual tracking algorithm (MAPF). Features are computed using the information
available from the output of the tracking system and finally actions are classi-
fied using them. Experimental validationwasperformedonrealvideosofhuman
actions and has shown that features extracted from a bounding box are good
enough to recognize different human actions.
MAPF is a very fast algorithm as well as the proposed feature extraction
method, which requires very little computation. Finally we use a SVM classi-
fier which has been proven to be very ecient. Overall, our action recognition
Search WWH ::
Custom Search