Machine learning for behavior classification of passively tracked vessels

Abstract

Situational awareness is a constant challenge for autonomous underwater vehicles (AUVs) due to limited communication to the surface, navigation drift, and a need to operate in busy areas around other vessels. One piece of information that would be valuable to both keep AUVs safe and inform autonomous monitoring missions is the location and behavior of nearby surface vessels. Passive acoustic data collected on a hydrophone array and processed on an on-board computer can provide bearing and time-to-intercept (TTI). This information can be used to classify overall boat behavior and inform AUV response: for example, investigating an area a boat has circled or avoiding an approaching vessel for vehicle safety reasons. Simulation studies were used to characterize trajectories for simple vehicle behaviors based on the bearing and TTI data produced by an existing passive tracking system. A classifier based on K-nearest-neighbor with dynamic time warping as a distance metric was used to classify simulation data. The simulation-based classifier was applied to classify experimental tracking data on boats completing different types of behaviors using bearing/TTI from dock-based and AUV-based hydrophone arrays.