Development of a human activity recognition component

Understanding actions and interactions of humans from the RGB-D sensor input can significantly improve cognitive functions of robots and help safely and smoothly incorporate them in the world of humans. Human activity recognition component will thus be a significant addition to the functionality of RoboComp. Human activity recognition in the video is an interesting and challenging task and recent research shows that there are different ways to address the spatial properties of human actions and their temporal dynamics. In the course of this project I plan to start with the simpler LSTM architectures and iteratively test and improve the model to achieve state-of-the art results on the selected datasets with the final goal of providing ready-to-use RoboComp components.