A unified framework for false alarm reduction using scene context from airborne sensors

This Small Business Innovation Research Phase I project will demonstrate the feasibility and effectiveness of utilizing scene and geometric context to improve target detection in aerial videos. The key innovation in this effort is a unified framework to place localized target detection in the context of the overall 3D scene, its constituents, and activities by modeling the interdependence of targets, scene elements, scene and sensor geometry, and target-movement patterns. In addition, enabling technologies will be developed to extract relevant scene and geometric context directly from the scene observables. These include extraction of scene elements (such as, roads, vegetation, buildings), scene geometric properties (e.g., rough surface orientations, relative depths, parallax motion field), and models of spatio-temporal properties of the targets in the scene (e.g., positions, scales, traffic patterns, etc.). Extraction of context directly from videos will ensure the independence of proposed architecture from prior knowledge regarding the scene. The proposed framework will be able to ingest the output of any target detector and apply contextual information to reject detections inconsistent with the extracted context model. The Phase I effort will include: development of enabling algorithms, implementation of the framework, demonstration of proof of concept, and quantitative evaluation of the proposed technologies.