An Intelligent Action Predictor from Video Using Deep Learning
PM Naieera, PD Anu, Mohammed Hikmat Sadiq · 2018 International Conference on Emerging Trends and Innovations In Engineering And Technological Research (ICETIETR) · 2018
The proposed system is an action recognizer that can effectively predict the type of action being performed from video. Being trained on a set of categories of actions called KTH dataset the system can actually predict the actions in those categories with a significant accuracy. Benefitting from deep learning the system is completely extendable to predict higher range of actions through training on larger datasets like HMDB and UCF101. The system uses a three stream Convolutional network comprising of a spatial, temporal and a movement stream, instead of traditional neural networks, that takes care of the frame characteristics gracefully.