Incremental behavior modeling and suspicious activity detection | |
| Author | Kan Ouivirach |
| Call Number | AIT Diss. no.CS-13-09 |
| Subject(s) | Signal detection Electronic surveillance |
| Note | A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Computer Science, School of Engineering and Technology, School of Engineering and Technology |
| Publisher | Asian Institute of Technology |
| Abstract | Video surveillance systems have become widespread tools for monitoring and law enforcement in public areas. Due to increasing crime, video surveillance systems are being deployed in more and more places. Video surveillance systems are needed to help security personnel prevent and respond to criminal activity in a timely fashion. However, human monitoring becomes increasingly expensive and ineffective as the amount of video data increases. Also, manual review of all video footage is time-intensive and error-prone. Automated anomaly detection can help to improve the effectiveness of human observers by separating the video stream into a sequence of “normal” and “unusual” events. However, much of the existing work has many limitations in this direction. Oftentimes, separate models for each distinct a priori known class of “normal” behavior are assumed. This could lead to the problem of typical behavior evolving and becoming more diverse to the point that the false alarm rates increase. The naive solution would be to retain all of the observation data and retrain the system periodically, but this requires storing all of the incoming data, requiring too much disk space. In this dissertation, we therefore propose and evaluate an efficient method for incremental automatic identification of suspicious behavior in video surveillance data. First, we develop a blob extraction method that segments blobs and an appearance-based blob tracking method that uses a forward-backward overlap method and color coherence vectors (CCVs) to maintain identity through blob merging and splitting cases. Second, we propose a new method for detecting shadows using a simple maximum likelihood formulation based on HSV color information. We find that the method outperforms standard shadow detection methods on three different real-world video surveillance data sets. Third, we propose a new method for clustering human behaviors in the context of video surveillance that is suitable for bootstrapping an anomaly detection module for intelligent video surveillance systems. We show that the method is extremely effective in separating anomalous from typical behaviors on real-world testbed video surveillance data. Fourth, we propose a semi-supervised batch anomaly detection method that self-calibrates itself from a small bootstrap set in which each bootstrap sequence is manually labeled as normal or suspicious by a human operator. Our method proves extremely effective, with a very low false alarm rate at a 100% hit rate. Finally, we propose an effective behavior modeling and suspicious activity detection method extending the batch method that incrementally learns scene-specific statistical models of human behavior without requiring storage of historical data. The incremental method’s false alarm rate drops below that of the batch method on the same data. In experimental evaluations on real-world testbed video surveillance data sets, the proposed methods prove to be practical and effective at inducing scene-specific statistical models useful for bringing suspicious behavior to the attention of human security personnel. |
| Year | 2013 |
| Type | Dissertation |
| School | School of Engineering and Technology (SET) |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Dailey, Matthew N.; |
| Examination Committee(s) | Clark, James J. ;Afzulpurka, Nitin V. ;Supavadee Aramvith ;Sanparith Marukatat; |
| Scholarship Donor(s) | Royal Thai Government; |
| Degree | Thesis (Ph.D.) - Asian Institute of Technology, 2013 |