Toward efficient collaborative classification for distributed video surveillance
J.B. Hampshire, Christopher Diehl · 2000
In this thesis, we propose a general strategy for automated video surveillance that relies on collaboration between the surveillance system and the user. Such collaboration enables the user to help the system incrementally acquire the necessary context for truly robust surveillance. The success of this strategy is dependent on the ability of the system to identify novel instances of known or unknown classes that it does not understand. This, in turn, allows the user to focus only on the observations with the highest uncertainty that require interpretation. Designing a real-time classification process that supports novelty detection is nontrivial. The real-time constraint dictates computational simplicity, whereas novelty detection requires a high dimensional feature space to aid in discriminating between the known and unknown classes. The majority of this work focuses on the problem of simultaneously satisfying these conflicting constraints. We consider these issues in the context of a relevant surveillance task and evaluate the performance of the resulting classification process in the CMU Cyberscout distributed video surveillance system.