Multicamera Action Recognition with Canonical Correlation Analysis and Discriminative Sequence Classification - Foundations on Natural and Artificial Computation - page 497

and optical flow are computed. Vertical and horizontal planes of the optical flow
are split and blurred. A radial histogram is computed over each of the optical
flow planes and the shape. The three histograms are concatenated into 216-
d vector. Lastly, PCA reduction of the surrounding past and future vectors is
appended to finally generate a descriptor of D TRAN = 286 dimensions. Readers
are referred to [18] for more details.
In order to sped-up the CCA computation, PCA analysis of the descriptors
is performed, retaining only the 100 principal components. CCA is going to be
trained for latent dimensionality values of q =10 , 12 , 14 , 16 , 18 , 205. The number
of hidden states of the HCRF is going to be fixed to | H | =11 , 22.

5.2 Results and Discussion

Table 1 shows the results obtained by the proposed method. It can be seen that
the maximum accuracy is obtained for q = 20 dimensions and 22 hidden states.
There accuracy starts growing with the number of dimensions, to then decrease
and again increase to achieve teh best results. These phenomena gives an idea
of how dicult is to manually parametrize the proposed methods

Table 1. Results obtained for different sizes of latent dimensionality

#dims

10 12 14 16 18 20

|H|

11 80.78 81.69 81.38 81.08 78.68 81.38

22 80.78 81.19 83.78 83.18 82.58 85.59

Finally, table 2 compares the results of our method to others. While it im-
proves the results achieved by other methods, it is still far from the results
obtained by the methods based on 3D visual hulls.

Table 2. Comparison of the accuracy of our method to others

Method Accuracy Type

Srivastava et al. [15] 81.4 Decision-in Decision-out

Our 85.59 Feature-in Feature-out

Weinland et al. [21] 93.33 2D Feature-in 3D Feature-out

Peng et al. [13] 94.59 2D Feature-in 3D Feature-out

6 Conclusions

This work has shown a preliminary approach to the fusion of features for human
action recognition using a subspace learning technique. Feature descriptors ex-
tracted from different camera views have been projected into a common subspace
learned using Canonical Correlation Analysis. The action classification has been