Real Time Detection and Tracking of Dismounts from Aerial Surveillance Imagery

Video surveillance for large-format imagery from airborne sensors represents a rapidly expanding source of information for counter-insurgent and counter-terrorist operations. Automated techniques to locate, represent, and recognize pertinent dismounts from vast video archives are becoming critical. Dismounts are non-rigid objects undergoing unpredictable/non linear motion and move slowly compared to vehicles. In this SBIR Phase I effort, we propose to build a video tracking module which consists of tools to reliably detect and track large number of dismounts from aerial video sensors. Our effort integrates in house algorithms for characterization of sensor motion and slow-moving targets detection, an adaptive non-linear appearance-based non-rigid tracking algorithm, an efficient multiple hypothesis target management module, and an on-line tracking evaluation component which can work in the absence of ground truth data. Our approach allows a threaded implementation with the capability of scaling to multi-core computers.