Analysis of Deep Learning Architectures for Human Action Recognition

Given the advances done in Image Classification and the great success this field is obtaining, Video Classification is the natural follow-up, although this task has more difficulties. This thesis describes the results of the implementations of some of the most successful techniques and approaches and then, the results of their fusion in order to achieve better performance and accuracy. Firstly, the Single Stream Architecture has been proposed with an LSTM implementation. Then, to consider also the temporal behavior, the Two Streams Architecture has been reproposed with two LSTM subnetworks. Moreover, in order to consider long-range temporal information as well, the Temporal Segment Network main concepts have been implemented. The implementations exploit the potentiality of Python, programming language, and Keras, a Deep Learning Framework that provides high-level neural networks APIs. The experiments exploit the Optical Flow Estimation techniques, namely Farneback estimation and LiteFlowNet warped estimation. They have been performed on UCF101 Human Action Recognition dataset. The results obtained are consistent with the state of the art techniques.