Phase II Amount
$1,497,400
The proposed AUTO 3D architecture leverages the robust set of tools and components available through AWS to optimize speed to an operational prototype, and efficiency of resources. The CDB Productivity API is containerized to transform the sensor data as it was recognized (points, imagery raster, meshed data) and produces the appropriate OGC CDB layers. Potential solutions may use OGC CDB raster material data and imagery signatures to improve segmentation and then apply those material codes to the polygonal surfaces to improve the data for simulation applications. If the data can be edge matched via pattern recognition to existing imagery to transform it into the correct location on the earths surface, it will improve the geospatial accuracy of the source data. Once the data is in the right location then the data needs to be segmented to provide a good Digital Terrain Model or Digital Elevation Model, and the rest of the 3D features extracted into OGC CDB models. AI/ML algorithms will be used to train and then invoke these transformation procedures, reducing the need for manual intervention to pick tie points between the imagery and the vector data after enough tie points are established to transform the vector data to the imagery to correlate the data. Team TAC offers a SOF unique, containerized interface for the creation, maintenance and execution of analytics models and training data focused on the transformation of small tactical UAS meshed terrain to simulation and GEOINT ready data. This interface enables users to create analytics models based on their ideas to experiment within the AUTO 3D architecture. This interface enables users to produce training datasets to empower the analytics models with diverse and quality datasets. The results of the experiments visualize within the containerized interface for users to tune their ideas. A key component of this interface is a Knowledge Management Capability to associate and configuration manage analytics models to training datasets. Team TAK will deliver analytics models and training datasets for segmentation, object detection and data restoration. The segmentation model focuses on a segmenting classic 3D objects from a mesh. This segmentation model will decompose the mesh into individual 3D objects (i.e. classic M&S 3D models) and terrain. The segmentation and object detection analytics models work in tandem to identify unique objects (i.e. a building or tree) within a 2D plane. Once an object is detected within a 2D plane, the segmentation analytics model removes the mesh components within a 3D space into singular mesh constructs. The singular mesh constructs are restored to address the oddities of all tactical UAS meshed terrain such as holes in the mesh, unnecessary mesh triangles and non-orthogonal planes (i.e. rippled instead of straight walls).
