Multi-camera security networks enable security personnel to monitor facilities and critical infrastructure for relevant activities.
However, the volume of streaming video generated by these security networks far exceeds the number of analysts needed for effective monitoring. Furthermore, activities and objects of interest to the defense and intelligence community are sufficiently rare that suitably large training sets are not available. As a result, performance for robust activity detection has not yet been realized for large-scale multi-camera networks. STR along with our academic and industry partners are developing a system for robust activity detection that unifies vision, language and geometric modeling to address this performance gap.