Machine Learning for Continuous Human Action Recognition

In this term project, we consider the problem of automatic recognition of continuous human activity. Our source data are short videos with RGB and depth information of seven predefined categories of human action as well as long videos that contain a series of continuous actions. By extracting frame-level features that represent each action and are invariant to small environmental noise, we train the model with part of our data and test it with the remaining data and then compare the precision of SVM and Order Representation model with different choice of features.