Multi-target Bayesian Tracking Phase II - 24-072

In response to SBIR topic N241-013 Metron has developed MultiLRT, a sequential Monte Carlo (SMC) Probability Hypothesis Density (PHD) filter to estimate the time-varying number of targets and their states jointly.  The “Multi” references the multi-target nature of the PHD approach, as well as its application to multistatics.  MultiLRT is envisioned as man and machine teaming system that integrates human intelligence, creativity, and adaptability with machine speed, precision, and computational power.  MultiLRT is designed to reduce operator workload by observing snippet data, developing hypotheses on the match-filter returns, initiating tracks supported by data over multiple pings, and maintaining custody of high-SNR or feature rich tracks.  In favorable environmental conditions, we demonstrate that MultiLRT will track ships and submarines that present with medium-to-high SNR, Doppler, and other features.  In less than favorable conditions, the MultiLRT automated tracker will maintain custody of higher-SNR observable tracks while providing a scoring mechanism to promote weaker contacts for investigation.  This novel scoring mechanism is based on the PHD surface, a principled multitarget version of the traditional LRT surface.   During Phase I, we successfully developed and demonstrated a Java-based Bayesian SMC PHD filter-based algorithm that employs an SMC-based implementation of a PHD filter that can manage non-linear measurements, complex performance models, and arbitrary contact features.  It produces statistical track probability particles or “tracklets” as a basis for track extraction.  Particle histories are used to compute a Jaccard Similarity Coefficient between pairs of contacts.  The similarity coefficient computed between every pair of contacts provides a basis for a hierarchical clustering from which sets of contacts that share track history can be extracted and presented to the operator.  The tracklets formed by our SMC PHD implementation also provide a way of retrospectively updating contacts scores to help initiate tracks with fewer contacts.  In our analysis we examine four different operational use cases of varying complexity and one demonstration of a PHD based scoring function.  The results clearly demonstrate a contact tracking capability that initiates and tracks over 20 multistatic contacts simultaneously without human intervention.  MultiLRT has also demonstrated the potential to track weak SNR contacts with update rates as low as 45 minutes.  Additional tuning with fielded operational contact screeners like the Support Vector Machine (SVM) scores will further improve target detection and clutter suppression capability.  A Java-based approach was selected during Phase I development to streamline integration into the Mighty Orion test environment and ultimately simplify transition into the P-8A Tactical Open Mission System (TOMS).