Illinois Rocstar proposes the development of AutoTrackML, a software for automatic target classification based on novel machine learning algorithms. Tactical platforms are challenged with the task of tracking multiple, closely spaced, targets with complex, unknown dynamics in near real-time from noisy sensor measurements. Performing calculations to compute these dynamics on tactical platforms is infeasible. To solve this, advanced machine learning models can be trained as an up-front cost to extract target dynamics from high-dimensional target data. The AutoTrackML software will produce automatic target classifiers based on the mrDMD (multi-resolution dynamic mode decomposition) machine learning technique. This method is well-suited to target classification due to its ability to extract multi-scale, transient dynamics in a fast, equation-free manner. The mrDMD classifier will be trained to embed physics based on realistic target data into the model. This model, combined with sparsely sampled sensor measurements, will allow for future state prediction of tracked targets. The technical feasibility of a mrDMD-based target classifier will be investigated during Phase I, with testing performed within the open-source tracking and state estimation library Stone Soup. In Phase II, the algorithms and software of AutoTrackML will be matured and deployed on tactical sensor platforms of interest to NAVAIR and the Defense Counterintelligence and Security Agency.
