Dismount Tracking in Urban Scenes

This SBIR Phase I project will demonstrate the feasibility and effectiveness of novel statistical models and deferred inference based techniques for robust detection and tracking of small targets in low spatial-resolution and frame-rate aerial videos. The key innovations in this effort include i) Bernoulli variable based statistical target detection framework that detects small targets by jointly modeling spatio-temporal properties of scene background and sensor geometry, ii) efficient multi-frame multi-target tracking algorithm that explicitly models common dismount tracking scenarios such as, track initiations and terminations, occlusions, target interactions, grouping behaviors, and noisy detections in a single optimization framework, and iii) forensic analysis system that enables the analysts to nominate a target of interest and efficiently determine its trajectory, the targets and objects that it interacts with, the trajectories of interacting targets, and the locations and times of interactions. The project will also benefit from ObjectVideo’s ongoing research activities on context based reasoning for false alarm reduction, context extraction, parallax modeling, and target detection in compressed aerial videos. The Phase I effort will include: development of proposed detection and tracking algorithms to enable the proposed forensic analysis system, demonstration of proof of concept, and quantitative evaluation of the proposed technologies.