Hybrid Learning and Model-Based Approach To Performance Prediction of Feature Aided Trackers

ABSTRACT: Generating track data from wide area motion imagery is an important first step in many exploitation tasks including high value target tracking, activity-based threat detection, and adversary network analysis. Tracker development to date has made significant advances, but there has been limited focus on tracker performance modeling and we need such a model to enable fusion with other sources of object detections and tracks, to establish our confidence in derived analysis products, and to quantify the value of additional collections or allocation of human resources to resolve tracker uncertainties. This program will develop a feature-aided tracking and performance modeling to dynamically compute measures of track performance, particularly distributions on track kinematic, association, and continuity, on a track-by-track basis, enabling users of that track information, whether human or automated, to perform the functions above. The performance model will be modular, enabling integration with both the baseline tracker we use for testing as well as other trackers. The tracker and performance model include on-line learning as an essential element to adapt context, object state, and performance model parameters over time. BENEFIT: The benefit of this program will be improved performance of video trackers and of downstream applications that leverage video tracks, and improved human, computational, and sensor resource allocation.