Physics-based Artificial INtelligence for Tracking and Classification using Neural Networks

To address the Navy’s need for an improved and more robust automatic target classifier, Intellisense Systems, Inc. proposes to develop a new Physics-based Artificial Intelligence for Tracking and Classification using Neural Networks (PAINTCNN), based on a novel combination of deep learning with neural networks, physics-based constraints, and tracking algorithms. To improve the process of performing automatic target classification, PAINTCNN uses preprocessing images to account for target and camera motion. Specifically, it takes advantage of ensemble learning and physics-based artificial intelligence (AI) modules, self-supervised few-shot classification, multispectral imaging constraints on colors, spatial-temporal multi-object tracking, and object size estimation. As a result, PAINTCNN offers improvements in combining new mathematical tools, physics-based AI modules, scaled and computer models, and sparse observational data, which directly address the platform requirements for improved accuracy, identification, and classification of complex or subtle dynamics by leveraging advanced mathematical and machine learning tools. In Phase I, Intellisense will demonstrate the feasibility of PAINTCNN by designing and developing a plan for implementing physics-based machine learning using sparsely sampled and noisy scaled laboratory data, demonstrating feasibility of a sufficiently robust system to handle—and complex enough to leverage—spatial and temporal coupling and dynamic motion. The Phase I effort will include prototype plans to be developed under Phase II. In Phase II, Intellisense plans to develop a machine-learning classification algorithm for multiple targets with separate quarantined targets that can be any class with spatial and temporal dynamics. We will build multiple well-trained networks including a Long Short Term Memory Recurrent Neural Network (LSTM RNN), a Support Vector Machine (SVM) classifier, and a Convolutional Neural Network (CNN) using physics-based hidden layers and scale model representations. These networks will test and demonstrate the extent to which sparsely and/or noisy data from the quarantined target can be incorporated into the existing classifier. These networks will also test the extent to which a trained hidden physics layer or a physics-based AI module can produce representative data that matches existing computer or scaled model data. We will also demonstrate the ability to generate data or a model that is robust against a well-trained SVM or CNN classifier and demonstrate the performance of the developed algorithm, possibly tested on an approved data set for validation.