Foveated Video Object Recognition

Low resolution quality of operational data, size of objects of interest, view occlusions, and crowded scenes degrade the performance of state-of-art region saliency and object recognition approaches when applied to overhead sensor data. We have developed a technology to automatically detect and recognize multitude of objects of potential interest providing an object recognition decision with a high level of confidence, and which can effectively and efficiently perform in real to near-real time on medium to high end desk top computer. Our technology employs state-of-art deep learning and bio-inspired methods to efficiently and effectively detect small objects in overhead noisy (crowded occluded) videos, where state-of-art models fail. The technology comes with the suite of the support modules that enhance user interaction and modeling. We propose to advance the Foveated Video Object Detection and Recognition technology into a suite of libraries and executables to meet the form factor and be integrated into PMA-281 projects, Common Control System (CCS) and Digital Camera Receiving Station (DCRS).