SBIR-STTR Award

Computational vision models (CVM) of the pre-cognitive, retino-cortical "front end" stages of human vision have demonstrated successfully explained visual performance for simple visual tasks with simple stimuli. However, extending these models to predict or simulate human visual performance for complex involving target discrimination such as discriminating, a tank from a truck, have not been very successful. Discrimination tasks involve cognition. Objects are examined and categorized based on prior knowledge of target structures and features. Expectations regarding what objects are present and where they will be found also influence visual interpretation. Successful modeling of this cognitive "back end" of vision requires a model of the prior knowledge of the visual appearance and structure of the scene. The proposed Phase I project will develop and demonstrate a "back-end" cognition model for visual target discrimination, The model will use a structural and spatial decomposition of the target into component parts and features as the basis for the model target categories. The features will actually be composed of the spatial filters used by the human visual system. In Phase II this model will be reviewed, refined, fully automated, calibrated, published for peer review, and demonstrated for a variety of military and commercial applications. This technology will reduce the time and cost of military and civilian vehicle design involving (1) camouflage, concealment and deception technologies, (2) automotive conspicuity enhancements crash avoidance, (3) imaging electrooptical sensor Systems for both target acquisition and enhanced driver's vision, and (4) visual identification friend or foe (IFF).

The objective of the project is to develop computerized image analysts tools for use in screening and evaluating vehicle design concepts. This project will implement, test, refine and calibrate predictive models of human observer performance in visual target discrimination. Models will be developed to address three aspects of visual discrimination: (1) segregating the target from its surroundings based on luminance, color and texture gradients, (2) categorizing simple targets based on the similarity of their exterior shape to members of a set of ideal or iconic forms, and (3) categorizing compound tar-gets based on the logical and spatial relationships among the individually recognizable simple component parts. Each subsequent model will build on the results of the preceding model. The models will interface with the existing TARDEC Visual Perception Model (VPM) of retinal-cortical visual processing. The VPM will be the "front-end" to the "back-end" visual discrimination models. The models will be iteratively tested and refined, before being calibrated for operational tasks. Developmental testing will focus on predicting human performance in character recognition. The eye charts used in character recognition testing will be created by applying selected image transformations and degradations in order to ensure that the model is accurate over a known range of target signature and back-ground clutter parameters. The models will then be calibrated to perception test data for military target acquisition and civilian driving tasks. Calibration data will come from defense vehicle technology testing and commercial automotive safety testing, with the cooperation of potential downstream Government and industry users. The participation of potential downstream users will help ensure commercial viability of the final products. Product: computer-aided design and analysis tools. Military applications: screen and evaluate camouflage, concealment and deception, and image enhancement technologies. Commercial applications: screen and evaluate crash-avoidance countermeasures, driver's visibility and enhanced vision systems. Dual-use applications: evaluate console displays and symbols.