Realistic UUV Data Transformation Tool

Undersea target recognition from sensor systems onboard unmanned underwater vehicles (UUVs) play a critical role in the US Naval strategies and mission capabilities. Machine Learning provides a game-changing opportunity for improved Automated Target Recognition (ATR), but current attempts remain limited due to a lack of adequate training data. ML-based ATR algorithms are statistics-based systems; with model parameters that are ultimately fit to the available training data. As with any statistics-based model, performance degradation occurs when the models are forced to extrapolate outside the bounds of their training data. Larger and more curated training datasets, with a focus on increased environmental and target variability, are required to increase performance of the next-generation ATR Machine Learning models. A new UUV data transformation tool, able to automatically process UUV sensor data from one modality, such as side scan sonar, and synthesize data representing same target and environment for different sensor modalities would provide a game-changing capability for ATR Machine Learning progression. A physics-based synthesis, with the additional capability to modify and adapt the environment and targets, such as varying target burial or adding surrounding clutter, provides an opportunity to increase the overall amount of training data by orders of magnitude and generate data for mission specific training. Makai proposes to develop an UUV Data Transformation Tool using automated data processing methods for high-fidelity three-dimensional modeling and state-of-the-art UUV and sonar simulation. ???????