Nachricht

Am Mittwoch, den 13. Januar 2016, hält Frau Biljana Jevtic um 16:30 Uhr im Raum P 1.4.17 einen Vortrag über ihre Masterarbeit mit dem Titel:

3D Convolutional Neural Networks for Human Action Recognition

Abstract:

In the recent years human action recognition has become a growing field of research. Most of the work, done so far, has focused on recognition of actions from videos taken in controlled environment. Assumptions, such as no background clutter or occlusions are usually made. Nevertheless, for real applications like video surveillance, these assumptions are not valid. First approach in this field was based on the feature extraction from the raw input data and then classification based on those features . In this approach, feature extraction is based on different methods, such as Cuboid detection, HOG/HOF, SIFT, etc. The main drawback of the methods belonging to this approach is that they are problem-dependent, since optimal features for one action class may not be appropriate for another one. To overcome this problem, the so called deep learning models have been used. The model proven to give good results in action recognition is the convolutional neural network (CNN) model. CNNs have recently be used for the human action detection. They work generally based on five channels of information: intensity, gradient in x and y directions and optical flow in x and y directions. In this master thesis two main goals are pursued: first implementation of the recent CNN based algorithms based on the five channels of information, and second investigating of applying different channels of information.