Date of Award
Doctor of Philosophy (PhD)
Graduate School of Computer and Information Sciences
James D. Cannady
Gregory E. Simco
This research produced a Predictive Anomaly Detector (PAD). It is an adaptive prediction-based approach to detecting unexpected events in data streams drawn from staring continuous-dwell sensors. The underlying technology is spectrum independent and does not depend on correlated data (neither temporal nor spatial) to achieve improved detection and extraction in highly robust environments. ("robust environment" refers to the data stream's control law being variable and the spectral content covering a wide range of wavelengths.)
The resulting approach uses a network of simple building-block equations (basis functions) to predict the non-event data and thereby present subtle sub-streams to a detection model as potential events of interest. The prediction model is automatically created from sequential observations of the data stream. Once model construction is complete, it continues to evolve as new samples arrive. Each sample value that is sufficiently different from the model's predicted value is postulated as an unexpected event. A subsequent detection model uses a set of rules to confirm unexpected events while ignoring outliers. Intruder detection in robust video scenes is the main focus, although one demonstration achieved voice detection in a noisy audio signal. These demonstrations are coupled to a concept of operations that emphasizes the spectrum-independence of this approach and its integration with other processing requirements such as target recognition and tracking. Primary benefits delivered by this work include the ability to process large data volumes for obscured or buried information within highly active environments. The fully automated nature of this technique helps mitigate manning shortfalls typically associated with sorting through large volumes of surveillance data using trained analysts. This approach enables an organization to perform automated cueing for these analysts so that they spend less time examining data where nothing of interest exists. This maximizes the value of skilled personnel by using them to assess data with true potential. In this way, larger data volumes can be processed in a shorter period of time leading to a higher likelihood that important events and signals will be found, analyzed, and acted upon.
Peter G. Raeth. 2003. Finding Unexpected Events in Staring Continuous-Dwell Sensor Data Streams Via Adaptive Prediction. Doctoral dissertation. Nova Southeastern University. Retrieved from NSUWorks, Graduate School of Computer and Information Sciences. (785)