Arthur Goshtasby (Advisor), David Miller (Committee Member), Mateen Rizki (Committee Member), Thomas Wischgoll (Committee Member)
Doctor of Philosophy (PhD)
Human action recognition is used to automatically detect and recognize actions per- formed by humans in a video. Applications include visual surveillance, human-computer interaction, and robot intelligence, to name a few. An example of a surveillance application is a system that monitors a large public area, such as an airport, for suspicious activity. In human-machine interaction, computers may be controlled by simple human actions. For example, the motion of an arm may instruct the computer to rotate a 3-D model that is being displayed. Human action recognition is also an important capability of intelligent robots that interact with humans.
General approaches to human action recognition fall under two categories: those that are based on tracking and those that do not use tracking. Approaches that do not use tracking often cannot recognize complex motions where movement of different parts of the body is important. Tracking-based approaches that use motion of different parts of the body are generally more powerful but are computationally more expensive, making them inappropriate for applications that require real-time responses.
We propose a new approach to human action recognition that is able to learn various human actions and later recognize them in an efficient manner. In this approach, motion trajectories are formed by tracking one or more key points on the human body. In particular, points on the hands and feet are tracked. A curve is fitted to each motion trajectory to smooth noise and to form a continuous and differentiable curve. A motion curve is then segmented into "basic motion" segments by detecting peak curvature points. To recognize an observed basic motion, a vector of curve features describing the motion is created, the vector is projected to the eigenspace created during PCA training, and the action most similar to a learned action is identified using the k-nearest neighbor decision rule.
The proposed approach simplifies action recognition by requiring that only a small number of points on a subject's body be tracked. It is shown that the motion curves obtained by tracking a small number of points are sufficient to recognize various human actions with a high degree of accuracy.
Furthermore, the proposed approach can improve the recognition power of other ap- proaches by recognizing detailed basic motions, such as foot steps, while introducing ef- ficient tracking and recognition compared to previous approaches. Recognition of basic motions allows a high-level recognizer to recognize more complex or composite actions by using the proposed system as a low-level recognizer.
Contributions of this work include reducing each video frame to a few key points on the subject's body, using curve fitting to smooth trajectory data and provide reliable seg- mentation of the motion, and efficient recognition of basic motions using PCA.
Department or Program
Department of Computer Science and Engineering
Year Degree Awarded
Copyright 2012, all rights reserved. This open access ETD is published by Wright State University and OhioLINK.