FA-PHD: A budgeted Focus-of-Attention framework for robust visual multi-target tracking

Multi-target tracking (MTT) in dynamic scenes increasingly relies on deep detectors, yet effectively fusing their intermittent outputs with Bayesian filtering remains a challenge. Existing RFS fusion schemes typically rely on globally uniform priors, unconstrained measurement-driven birth, or adaptive parameter tuning based on scene-level statistics. However, these scene-wide approaches often fail to resolve localized detection-prediction conflicts as detector errors exhibit strong spatial structure, resulting in unstable tracking especially under occlusion and detection noise. Inspired by the resource-limited nature of human visual focus, we propose a Focus-of-Attention (FA) birth framework that transforms passive initiation into an active, budget-aware Deep-Bayesian fusion process. This framework concentrates birth intensity on regions of potential new targets and suspected missed detections, while structurally redistributing a fixed birth budget to prevent cardinality drift, thereby enabling immediate target initiation and strong resilience to detection uncertainty. The proposed framework is instantiated within the Gaussian Mixture PHD filter to establish the FA-PHD algorithm, incorporating an Anchor-Guided Structure-Aware Extraction (AGSE) module to suppress redundancy and stabilize state reporting. Extensive experiments on complex traffic and crowd benchmarks demonstrate that FA-PHD outperforms representative RFS baselines in tracking accuracy while retaining practical runtime efficiency, providing an effective solution for robust tracking under detection uncertainty.