Robust object detection using context and component information

点击翻译按钮获取中文摘要

Aerial/space remote sensors acquire overwhelming amount of data at the present time for opto-electronic surveillance with multi-spectrum, multi-temporal, georeference data and significantly increased spatial, spectra! and radiometric resolutions. High-resolution aerial images permit to extract 3D information on the grounded objects. Manually processing those data for image registration and fusion, environmental change identification and object detection could be time-consuming and inconsistent for the large-scale image bank. The similar problematic exists in the civil digital libraries, which contain huge amount of visual data in the multimedia and database, thanks to rapid growth in the multimedia acquisition, storage, transmission, visualization, and interaction technologies.The proposal consists of two related projects: context-driven surveillance and component-based 3D object detection, addressing two of the basic problems in aerial surveillance: where to search and how to detect the objects. Context-driven focusing of attention will be studied using the new beamlet transform, which searches for object-of-interest and the neural networks and support vecotr machines for context operations. Detection of 3D object will be studied in terms of invariant feature and component feature extraction and of height information estimation using simultaneous localization and map building techniques.The proposed research will deliver internal reports, publications and effective software to the three partners: Defense Research and Development Canada (DRDC) Valcartier, Optronic Surveillance Data Exploitation Group and Lockheed-Martin Canada (LM). The objective of the proposed research is to acquire, develop and generate new knowledge and expertise on the state-of-the art technologies in the aerial/space electro-optic image surveillance and 3D object detection.