Action recognition using combined local features

Thispaper presents a new algorithm for recognition of actions based onlocal space-time features. The algorithm resulted from intensiveresearch of classification and feature extraction and it is anextension of the previously known ones. The most importantachievement is that it is shown that through carefully selectedcombination of space-time features leads into better precision ofrecognition on some events comparing to the state-of-the-artalgorithms while it is comparable on all other events. The paperdescribes the algorithm, its main features and improvements,demonstrates the achieved result, and draws conclusions.