This thesis develops an approach for detecting behavioral anomalies using tracks of pedestrians, including specified threat tracks. The application area is installation security with focus on monitoring the entrances of these installations. The approach specifically allows operator interaction to specify threats and to interactively adjust the system parameters depending on the context of the situation. This research has discovered physically meaningful features that are developed and organized in a manner so that features can be systematically added or deleted depending on the situation and operator preference. The features can be used with standard classifiers such as the one class support vector machine that is used in this research. The one class support vector machine is very stable for this application and provides significant insight into the nature of its decision boundary. Its stability and ease of system use stems from a unique automatic tuning approach that is computationally efficient and compares favorable with competing approaches. This automatic tuning approach is believed to be novel and was developed as part of this research. Results are provided using both measured and synthetic data.
Library of Congress Subject Headings
Machine learning; Automatic classification; Pattern recognition systems; Image analysis--Mathematics
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Zelnio, Holly, "An Approach to detecting crowd anomalies for entrance and checkpoint security" (2012). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus