Recognizing Human Actions by Using Spatio-temporal Motion Descriptors - Advanced Concepts for Intelligent Vision Systems - page 371

Fig. 5. Example video frames from the negative samples: hand selected video sequences
(top); randomly cropped from the recordings (bottom)
0.75 aspect ratio) at random spatial and temporal positions. This dataset con-
tains 63 sequences. The two negative sample sets currently contain 130 video
sequences. Fig. 5 demonstrates some frames from the two negative samples.
5 Experimental Results
In our experiments we tested the two descriptors presented in Sec. 3.1 and
Sec. 3.2, and both quantization arrangements (see Fig. 1) were evaluated with
4, 6, and 8 bins. Moreover, we also included in our tests the publicly available
implementation of the HOG3D descriptor[9] using the default settings.
In our first experiment we used the drinking dataset presented in Sec. 4.1 to
evaluate the performance of our methods. Two tests were performed with differ-
ent size of training data. In the first test for training we used 10 samples from
the positive and 11 from the negative dataset, while in case of the second test
we increased the number of training samples to 17 (positive) and 20 (negative).
The rest of the datasets were used for testing, and the trained SVM was used
to recognize the action. Table 1 presents the confusion matrix of the recognition
results of the first test with the smaller training set. Here we show the best
results obtained by each descriptor, which were achieved in both cases by using
the arrangement presented in Fig. 1b with 6 bins. Due to the larger training sets
used in the second test the number of False Negatives decreased to FN=0, while
the number of False Positives changed to FP=1 for HDG and HFD .Incaseof
the HOG3D we obtained FN=0 and FP=1.
We used our stand up dataset (see Sec. 4.2) in the second experiment. Again,
two tests were performed with training data of different size. In the first test we
used 24 samples from the positive and 36 from the negative dataset for training,
while in case of the second test we increased the number of training samples to
41 (positive) and 87 (negative). Table 2 summarizes the recognition results of