Abstract:
Human action recognition is an interesting area of research that has found applications in security surveillance systems, robotics, human-computer interaction and so on. Human actions can be classified into usual (mundane) events and unusual (peculiar) events. The traditional supervised learning models that discriminate between classes are helpful in classifying mundane actions of which the data is available during training. But in case of unusual events, we do not generally possess the example data during training. This now becomes a problem of zero-shot learning. In this study, I explore generative models to produce instances of peculiar action events with the help of semantic meaning related to the action classes. The data from different places at AIT has been combined with benchmark UCF101 dataset to carry out the experiment. I have achieved an accuracy of 96.7\% on the usual event classification and 60.94\% on unusual event classification tasks during test time.