Robust Maritime Target Recognition

The proliferation of unmanned aerial systems provides an excellent opportunity for robust, passive, maritime target recognition and classification.  These platforms can be outfitted with (1) high-quality camera systems, (2) high-performance compute resources, and (3) encrypted links for relaying tactical information to other nodes.RDRTec has already demonstrated the ability to perform maritime target classification from ISAR imagery at very long ranges (100’s of km) which not only separates combatants from commercial vessels with excellent accuracy but further is able to distinguish combatants at levels of detail even below fine naval class.  This Automatic Combat Vessel Classifier (ACVC) divides ships of the world into over 1500 types.  Classification is done using a sensor agnostic, physics-based classifier and a reference database populated with  all of the templates.  This approach requires no training and performance has been demonstrated on datasets from new sensors with no parameter tuning.ACVC ’s profile feature extraction will work with almost no modifications for EO/IR imagery taken from angles +/- 65 deg from broadside and grazing angles up to 15 deg as optical imagery is similar in presentation to the ISAR imagery which  ACVC currently processes.   ACVC feature extraction consists of identifying superstructure location and width at various heights from the deck line.  The ACVC database is populated with features for all 1500+ templates, and we have demonstrated that we can approximate the expected extracted features from these viewing angles correctly from a single ISAR profile view and perform classification based on this feature set.The predecessor of RDRTec’s ACVC system, MCA, has already been integrated with mission systems such as the USN’s Minotaur system and sensor resource managers systems.  It has been hosted on flight ready tactical processors.  Part of  ACVC’s design requirements has always been the ability to run in real-time on modest processor resources.  Currently we have a delivered configuration that can run in a single 3U Open-VPX slot while performing not only classification but also performs ISAR image formation starting from raw IQ data.RDRTec is currently advancing a hybrid physics-based / ML addition to ACVC for exploiting vibration and rotating aperture signatures and will take a similar approach in augmenting ACVC for use with EO/IR sensors across wide viewing angles. During Phase I we demonstrated successful optical image-based classification using the ACVC system with default ISAR settings. These efforts included a proof-of-concept vessel foreground extraction algorithm using custom developed MATLAB code and manual processes for creating input to ACVC. These were effective in confirming the feasibility of our approach.  In Phase II we will develop processes to provide a fully automated classification pipeline  that operates similar to the existing ISAR-based system.