SBIR-STTR Award

We propose to detect and identify moving objects in airborne imagery and full-motion video in unconstrained environments. Conventional techniques result in too many false positives; a new topology is needed to automatically detect and identify moving targets from a moving platform in airborne imagery. Our software is designed to analyze video imagery and deliver curve metadata that can be organized into an SQL database of relevant objects and regions of interest. This is accomplished utilizing scale- and angle- invariant Bezier curves. Effectively, we are taking image analysis out of the realm of pixels, and into more reliable mathematical models; we call this method Shape-Based Modeling Segmentation (SBMS). SBMS Bezier elements have a marked tendency to discover and "attach"? to orderly edges in images without training; this is a result of Vy's underlying order-recognition algorithm that rejects gradient edges that fail a test for continuity and differentiability. Groups of Bezier elements discovered in this way can be associated based on their spatial alignment, conformation to object models (such as ellipses and other closed contours), and photometric similarity; these discovery and filtration techniques can then be combined orthogonally

Artificial intelligence and deep learning systems used for image analysis have a significant limitation: there is no way to explain how their decisions are made. Additionally, deep learning systems are heavily dependent upon RGB color which is highly variable necessitating the need for thousands of training vectors. Vy has developed a powerful set of algorithms collectively called Shape Based Modeling Segmentation (SBMS.) We apply transparent and auditable mathematical models (Bézier curves and decision trees) to collect hard data from visual imagery that significantly increases the speed and accuracy of object recognition. The premise of this project is that by combining deep learning with SBMS ion a synergistic effect will result in improved performance in convergence speed (number of training vectors) and improved accuracy, as reflected in the precision of object classification and in correct detection of edge location. In Phase I of this project we achieved an order of magnitude improvement in convergence for three object classes (planes, trains, and cars) resulting in the system learning more quickly with fewer training examples. In Phase II, we intend to develop a generalized approach for many object classes that is suitable for commercial development.