| Abstract [eng] |
In this paper reviewing literature that deals with action recognition using various dimensional data (2, 3, 4, and n-dimensional) and different methods for forming feature vectors. Selected large-scale Human3.6M database with excluded postures feature vectors after literary analysis. For these data classification chosen TreeBagger, Neural network and LibSVM classifiers. To allow manipulation of large amounts of data adapted strategy such as large-scale data classification, vote, bundling rules of classification models, one class isolation from all the rest, reduction of recognizable actions. In the last stage for grouping data chosen model with the highest accuracy in overall recognition of all classes. In this case working with classified data, these data are grouped into certain sequences, deciding which class detection probability is the highest for each sequence. Reached 82,63 % overall recognition accuracy of 15 classes for which data were not used in the training process. |
