Density-aware reliability association for training-free FDTA enhancement in multi-object tracking

Abstract

Multi-object tracking (MOT) in crowded scenes is difficult because objects frequently occlude each other and share similar appearances, leading to identity confusion. End to-end trackers based on the From Detection to Association (FDTA) architecture reduce inter-object embedding errors but remain susceptible to ID Switches (IDSW) when ID confidence falls briefly during crowd traversal. This paper proposes a training-free density-aware association module that operates on top of a frozen FDTA model at inference time. Identity recovery activates only when both local density and frame-level detection count exceed fixed thresholds, preventing false merges in sparse regions. The method also includes a geometry-based offline tracklet linker that reconnects fragmented trajectories using bounding box constraints alone, with no learned components. On the DanceTrack validation set, the approach reduces IDSW by 13% and raises Higher Order Tracking Accuracy (HOTA) from 64.36 to 65.39, Association Accuracy (AssA) to 57.15, and ID F1 Score (IDF1) to 71.47. On the official test server, the method scores 72.01 HOTA and 78.21 IDF1