Transformer‑based multi‑target tracking (MTT) learns data association and state estimation jointly, yet MT3/Track‑MT3 style trackers repeatedly re‑encode measurement windows, incurring redundant computation. We introduce Causal Neural Set Filtering (CNSF), which encodes only the current measurements while preserving past evidence in a structured recursive track state. CNSF integrates three key mechanisms: exclusive Sinkhorn association that enforces soft one‑to‑one matching, association‑conditioned Kalman‑shaped updates with moment matching that propagate measurement‑induced uncertainty to the state distribution, and a recurrent Bernoulli lifecycle model with measurement‑driven birth that estimates target existence under missed detections and birth‑death transitions. On a held‑out simulated test set covering three motion regimes, CNSF reduces mean GOSPA and T‑GOSPA by 19.3% and 30.4% relative to Track‑MT3, uses 55.9% fewer parameters, and achieves a 3.76× speedup on single‑thread CPU inference. The implementation is open‑source at Code.
Review