SBIR-STTR Award

M-Vision Inc., along with the University of Michigan and Michigan State University, proposes to develop a new and novel solution to the problem of target detection and discrimination in a complex scene. This solution is obtained by innovatively combining Gabor wavelets and scale space inference. The result is a computational model of human perception that resembles the human visual system in its information gathering stages, its cognitive or information processing stages, and the manner in which these two stages interact with each other. At the heart of this model is a mechanism to learn and reason about images much like humans do. This is achieved by building a knowledge base about image features and constructing a Fuzzy rule book to map observed image features onto a target-background detectability metric. As a consequence, every image is interpreted in a dynamic fashion - features that are significant for that particular image are first obtained and then the importance of these for target-background detectability is assessed. In Phase I, a feasibility study is proposed, including the design of appropriate experiments and the establishment of a protocol to assess performance. The deliverables at the end of Phase I include the new Gabor wavelet based target signature metric, the associated human cognition based scale space inference engine, and the entire set of experiments conducted and results obtained. The proposed computational model of human perception has wide spread potential use among Department of Defense scientists and engineers in their survivability, signature analysis, and sensor characterization studies. Furthermore, the model also has much potential use among scientists and engineers in commercial automobile and other automotive-related industries in their conspicuity analysis, collision avoidance, all-weather driving assistance, and realistic driving assistance, and realistic driving simulators studies.

M-vision Inc., along with VisiTek inc., proposes to engage in a phase II study to design and uild a multi-spectral target acquisition and analysis system. Considerable attention will be paid to the system's usability, so that it would be of the 'plug-and-play' variety. The system design will be modular, so that the various sub-components of it can be inter-changed as additional sensors and algorithms become available. The system will comprise of a suite of sensors - visual, near-infrared, short-wave and long-wave wave infra-red sensors - and relevant accessories such as lenses, prisms, mounts etc. The sensor suite will Se combined along with a portable personal computer, and associated interface hardware and software. All the equipment used would be off-the-shelf, and readily available commercially. The images from the various sensors will be properly registered against each other, so that target signatures across multiple spectra can be easily interpreted. The role of polarization to enhance target signatures will be investigated thoroughly. A carget analysis software package will also be developed as part of this proposed study. This package will contain algori~hms to extract low-level target/background feacures, such as edges, texture, motion, etc. Higher-level algorithms to interpret these low-level features will be provided, using two different approaches - a fuzzy rule-book based approach and a neural network based approach. These higher level algorithms will be able to both learn as well predict the targets' conspicuity, and they will be easy to use. In the end this phase II study, it is expected that a prototype multi-spectral target cquisition and analysis system will result, which will then be made into a produce and marketed in phase III. The deliverables at the end of the proposed phase II study include, the complete multi-spectral target acquisition and analysis system, a final report providing details of the system (hardware and software) design and a users manual explaining how to use the system.

Keywords:
TARGET ACQUISITION MULTI-SPECTRAL SENSORS ONSPICUITY ANALYSIS FUZZY INFERENCE COMPUTATIONAL MODEL NE