Anomaly Detection for Data Reduction in an Unattended Ground Sensor (UGS) Field.

摘要

This report describes the design and implementation of a data reduction technique for video sensors that are part of a larger unattended ground sensor (UGS) network. The data reduction technique is based on anomaly detection in full-motion video and subsequent statistical analysis techniques that allow the system to identify abnormal or otherwise interesting behavior that acts as a notification trigger. These techniques have been integrated into an existing distributed sensor framework at the US Army Research Laboratory (ARL) that is based on the Open Standards for Unattended Sensors (OSUS), developed in collaboration with the Defense Intelligence Agency (DIA). Furthermore, the statistical technique applied to this problem does not require any training data, which are often impractical in battlefield environments. Instead, it operates by being bootstrapped on mission profiles and templates established by the user and supplemented by incremental and online learning algorithms. Overall, these techniques combine to provide an effective approach to monitoring large distributed sensor fields without network and operator overload.