TOYON RESEARCH CORPORATION — Department of Defense SBIR Phase I: AF161-138

Description

ABSTRACT: Toyon Research Corp. proposes to develop innovative cognitive processing algorithms to perform ultra-fast identification and classification (uFIC) of structural information directly from 3-dim LiDAR point-cloud data. The proposed algorithms are based on a model of human cognition, which involves coarsely processing data to identify regions of interest (ROI), and subsequently increasing the granularity of detection to resolve objects and patterns. Like the human visual cortex, our algorithmic cortex is composed of several layers, which filter and interpret local point-data correlations at different scales using simple geometric models. Multi-scale point and group associations are stored in efficient membership structures, which are built at each layer and refined with low complexity updates. The resulting system is highly tunable, via a number of objective- function parameters, and adaptable to various object geometries and signal modalities. Furthermore, the system is capable of real-time performance, with immediate extensions employing massively parallel GPU and distributed CPU architectures such as OpenMPI.; BENEFIT: The tool will provide an intuitive user interface for visualizing, searching, and annotating 3-dim LiDAR datasets in real-time. The prototype will be based on a geographic information system (GIS) framework with a graphical user interface, 3D environment models, terrain and overlay visualizations, import/export of standard data formats, and accompanying CAD tool to generate synthetic models for training. The recognition algorithms will consist of low-level geometry extraction codes coupled with a trainable software stack (modeled on human neural anatomy and deep learning architectures) that can exploit training data. The resulting tool will enable an analyst to load up a point-cloud file, automatically find geometric objects from a training set, and visualize them using in a scene using simple shading and wireframe models.