Intelligent, Modular Network for Detection and Tracking Military Threats

This proposal addresses the need for improved situational awareness specifically in terms of chemical hazard sensing and motion of adversary assets including people and materiel. The proposed project will develop and demonstrate signal processing and decision support tools for use with multimodal, remote sensors suitable for chemical hazard sensing as well as surveillance, tracking and targeting.áIn Phase I project N5 Sensors will develop and demonstrate the feasibility of configurable and modular algorithms and decision support tools to accommodate a variety of types and platforms operating remotely. The algorithms must be capable of highly efficient operation in terms of size, weight and power using new embedded microcontrollers and application specific integrated circuits. One challenge in multi-model datasets is the disparate spatial and temporal resolutions of each sensor modality (i.e. some modalities may have a different sample rate, response behavior, and spatial arrangement compared to other modalities). Additionally, any model designed to run on an edge device or resource-limited device will need to maximize efficiency. To tackle these challenges, N5 Sensors in collaboration with Professor Abhinav Shrivastava, University of Maryland (UMd) Department of Computer Science proposes the following technical approach for development of an accurate and efficient multimodal detection system and reporting system for military threats, both chemical and materiel.Convolutional Neural Networks (CNNs) are being successfully applied to a variety of sensing problems including chemical detection. The downside of the CNN approach is that it struggles with non-Euclidean datasets, meaning that the CNN is built for structured connections between nodes. Graphical Neural Networks (GNNs) however can have different number of connected nodes and they are unordered, more representative of fielded deployment where the nodes are scattered. A GNN can have different number of connected nodes and they are unordered, and this better represents the intended deployment where the nodes could be scattered. N5 has been using Graphical Neural Networks for data fusion and analysis in remote, wireless sensor networks currently deployed for wildfire and associated air quality monitoring.áThese multimodal sensing systems gather data from different sensor types in a detection node and fuse the data from the network consisting of multiple nodes. Graph Transformer Networks (GTNs) will also be evaluated, and they are able to handle missing data in the network more handily. The Phase I conceptual design will be initially assessed for interoperability based on standards provided by the Government. In Phase II the evaluation will be extended to include testing on representative hardware platforms comprised of diverse chemical sensor inputs (i.e., different sensor modalities) as well as geospatial datasets.