Poster: Advanced Feature Based Deep Learning For Intelligent Human Activity Recognition: An Approach Using Scene Context And Composition Of Sub Events

We investigate the problem of automatic action recognition and classification of videos, using deep learning techniques. Deep learning specializes on generating hierarchical features in spatial domain, but extending it to temporal axis, still remains an open problem. In our work, we explore how to optimally provide calculated low level motion features to the network, as the network itself is not able to capture temporal dependencies from raw input frames. We also experiment on the effect of providing static scene context information to the network, in the task of recognizing actions. Furthermore, we focus on how to compose sub events, in order to obtain a higher level semantic meaning for more complex events, by setting up a network of LSTM units on top of the system.