Hierarchical Nonparametric Tracking (HINT)

The development of full motion video sensors on board unmanned platforms is quickly enabling persistent surveillance of areas of military interest in dense urban environments. Although the high resolution and update rate of persistent sensors facilitates tracking ground targets, serious tracking challenges remain due to obscuration, closely spaced targets in traffic, and maneuvers. Recently developed non-parametric Bayesian statistical methods for stochastic dynamical systems provide an opportunity to improve tracking by predicting track performance in such environments by learning target signature and motion behavior on-the-fly from the sensor data. The methods also automatically generate statistics which can be fed directly into fusion, sensor management, and other track exploitation systems. The proposed effort will apply non-parametric Bayesian methods to predict performance in feature-aided video trackers. The main output of the effort will be a prototype MATLAB implementation of the track performance prediction algorithm embedded in a MATLAB implementation of a forensic video tracker. In addition, the effort will produce an evaluation of the algorithm"s performance with simulated data and CLIF video data. Although developed for video sensors, the technology will apply to other types of moving target sensors. BENEFIT: The research will produce automated algorithms to predict track performance for feature-aided trackers using video or other types of sensors. The algorithm learns target signatures and motion behaviors on the fly, and outputs statistics which are needed by track fusion systems, sensor management systems, network detection, and other exploitation systems. The technology will improve ground-target tracking for surveillance of moving targets (vehicles and dismounts) in complex urban environments for military, law enforcement, emergency management, and security monitoring applications.