Vision-Based Autonomous Sensor-Tasking in Uncertain Adversarial Environments

The key objectives of the proposed research was to build theoretical and computational foundations for developing efficient, robust and adaptive data-driven exploitation techniques and tools for automatic activity analysis in surveillance applications, and to incorporate dynamical systems, control theory, computer vision, machine learning and statistical techniques, through analytical and numerical methods, in the design of surveillance systems. The first goal is aimed at technology transition by creating and transitioning to Air Force generic tools that reduce analyst workload, enhance analysts situational awareness, and increase analyst efficiency and effectiveness in discovering and forecasting potential anomalous activities, and exploring hypotheses about those potential anomalous activities. Current autonomous sensor networks generate vast amounts of data while monitoring complex uncertain environments, provide limited actionable information, and are limited by the required number of human analysts. The aim of the second goal is to leverage dynamical systems and control theory to further optimize machine vision based surveillance systems such that they enable long term activity forecasting and early anomalous event detection provide desirable tradeoff between false alarms and missed detection and exhibit robust performance under varying environmental conditions and scene contexts. Current machine vision systems have limited ability to exploit context in the activity analysis, forecast activities, and analyze complex scenes with multiple interacting entities. Specific applications include autonomous aerial surveillance systems that cover broad areas of military operations, camera security systems that cover large crowded areas in urban environments, and large-scale wireless sensor networks that must minimize power consumption while providing actionable system state information.