In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
Fatima Zohra Daha
will defend her proposal
Dynamic Graphs for Abnormality Detection in Videos
AbstractIn the current era of digital technologies, the number of surveillance cameras used, both indoor and outdoor, keeps increasing. These cameras are continuously generating streaming data that can be analyzed to (i) understand the scene; (ii) monitor events that would be of concern for public safety; and (iii) detect anomalies. Over the past several years, considerable attempts have been made towards video analytics for anomaly detection in videos due to its ubiquitous applications in real-world scenarios, e.g. analytics for left object (baggage) detection or line (perimeter) crossing are common today. However, these tools require a large amount of resources (e.g. extensive human involvement) to initialize or define parameters for effective utilization. The anomaly detection task in intelligent video surveillance still remains very challenging. The main difficulties of this task lie in the diversity of possible events as well as their short duration. Motivated by our observation that motion information is crucial to good anomaly detection performance, we propose a two-stage semi-supervised learning approach based on dynamic graphs capable of detecting anomalies in videos. In this work, we investigate both offline and online methods. We model the scene as a graph using coherent motion detection and dominant motion extraction, then we track its evolution over time by monitoring the changes in the graph properties. We validate our approach on two publicly available anomaly detection datasets at different scales capturing real-world scenarios. Quantitative and qualitative results demonstrate the efficacy of the proposed method and its generalization ability.
Date: Wednesday, December 04, 2019
Time: 2:30 - 4:00 PM
Place: PGH 550
Advisor: Dr. Shishir K. Shah
Faculty, students, and the general public are invited.