Simulating actions for learning

Abstract

This paper presents a novel approach for generating new actions to learn supervised algorithms such as the Adaboost in the context of human action recognition. Indeed, the learning process requires a large amount and variety of data. Our motivation in this work is to reduce the dependency on public databases and allow learning with small sets of actions. We overcome the problem of nondiscriminatory action datasets for action recognition by enlarging a set of actions performed by different persons in different ways and captured by a Kinect. We present a way to enlarge the originally captured dataset from a Kinect device or from simply annotated data. This is done by combining the extrema of the action sequences into intervals, creating random points within them, and adding certain variables to discriminate the samples. These actions are learned and tested with a late fusion Adaboost using simple features and a strong classifier for each joint. Finally, a confidence coefficient is calculated and used as input of a higher level Adaboost classifier.