Track-Before-Detect Adaptive Birth Using Generic Observation Model Labeled Random Finite Sets

Labeled random finite set filters are a popular option for multi-target tracking in challenging high-clutter, low probability of detection scenarios. Recent advances have shown that the labeled random finite set formalism can be extended to the track-before-detect setting, but lack a thorough discussion of measurement-driven track initialization. This paper presents an approach for generating a labeled multi-Bernoulli prior density on target states from radar power measurements. The prior density is generated adaptively from radar measurements and the multi-object density of previously persisting targets via a thresholding procedure that assumes a constant backscatter model (i.e., Swerling 0), Gaussian point spread function, and complex Gaussian receiver noise. The threshold per cell is constructed using a fixed false alarm rate that accounts for the contribution of energy from existing targets in the scene. This procedure improves upon existing approaches by enabling tracks to be initialized in the proximity of existing tracks. The proposed approach is demonstrated within a generic observation model labeled random finite set filter over a challenging scenario containing multiple overlapping target returns and target spawning dynamics.