A Comparison of Indexing and Sequencing with Dynamic Time Warping for Recognition of Human Activities

Human activity recognition has been one of the most important yet challenging areas of research in the field of computer vision. Various vision-based human activity recognition applications have been the constant motivation behind this research. The goal of human activity recognition is to track and understand the behavior and activities of humans in real time using video sequences obtained from 2D or 3D cameras. Many algorithms have been researched and developed in activity recognition, using 2D cameras, achieving significant recognition rates. In recent years, the emergence of 3D cameras such as the Microsoft Kinect has been a breakthrough for the research on human activity recognition. The extra dimension added by the 3D cameras has enabled researchers to develop more reliable recognition methods that are invariant to a person’s orientation and position. In this thesis, we propose a robust skeletal based recognition framework capable of recognizing activities with inter class variations. We compare the sequence recognition method Recognition by Indexing and Sequencing (RISq) with the renowned algorithm Dynamic Time Warping (DTW) in recognition of a wide range of human activities. The comparison includes testing both the algorithms for activities performed at various speeds, timing and interference in the forms of additive noise and missing data. We capture activity sequences using a Microsoft Kinect and with the help of the software developments kits available for the Kinect, we extract the skeletal information of the human. The captured skeletal information is normalized so that it can be used as feature inputs for both the algorithms. Using these features, we train both the algorithms to recognize and classify an unknown activity sequence.