Semi Supervised Learning for Human Activity Recognition Using Depth Cameras

Abstract

Human action recognition is a very active research topic in computer vision and pattern recognition. Recently, it has shown a great potential for human action recognition using the 3D depth data captured by the promising RGB-D cameras, and particularly, the Microsoft Kinect which has made high resolution real-time depth cheaply available. Several features and descriptors have been proposed for depth based action recognition, and they have given high results when recognizing the actions, but one dilemma always exists, the labeled data given, which are manually set by humans. They are not enough to build the system, especially that the use of human action recognition is mainly for surveillance of people activities. In this paper, the paucity of labeled data is addressed, by the popular semi supervision machine learning technique "co-training", which makes full use of unlabeled samples of two different independent views. Through the experiments on two popular datasets (MSR Action 3D, and MSR DailyActivity 3D), we demonstrate that our proposed framework outperforms the state of art. It improves the accuracy up to 83% in case of MSR Action 3D, and up to 80% MSR DailyActivity 3D, using the same number of labeled samples.