Human Action Recognition in realistic scenarios - actioncliptrain00014.flv

Some of the action classes may be context-free which means that inclusion of the scene in which the action is performed will bring no extra information about type of the action class. This could be due to two different reasons: The first possible reason is vast diversity of the context in which the action can be performed; for example sit-down action could be done within very many different scene contexts. Second reason could be the fact that for the context-free actions the video frame is almost completely filled by the actor and there is not much part of the scene context visible; for example the kissing action is most likely videotaped in a close frame view and it is expected to contain not much part of the scene. Nonetheless, it is generally expected to have a strong correlation between type of the scene in which the action is performed and the action itself. For example sit-up action most likely happens in a bedroom scene context. Therefore, generally speaking, it is helpful to use scene features beside motion features in order to boost up accuracy of action recognition methods. In this project we will basically try to incorporate scene features with motion features and study the mutual effect of these two types of features.

In this project, I designed the GUI tool to read the videos and create the annotations on them. The annotations are defined with user inputs at some keyframes, while they are interpolated between these keyframes. It allows a fast annotation process, while being accurate and not too tedious.

Category:

Science & Technology

Tags:

• Human
• Action
• Recognition
• realistic
• scenarios
• gui
• image
• processing
• analysis
• c++
• opencv
• bounding
• box