PrismTrack: Perspective-Aware Multi-Cue Association for Robust Multi-Object Tracking

Abstract

Multi-Object Tracking (MOT) remains challenging due to object occlusion, complex motions, and detection unreliability in crowded scenarios. We propose an enhanced MOT framework integrating and optimizing state-of-the-art components, specifically Improved Detection Confidence Boost (IDCBoost) and Track-Perspective-Based Association (TPA). The confidence boosting process for recovering more potential detection candidates is refined by introducing Bbox-Based Distance (BBD), a deterministic metric that compensates for the instability of conventional Mahalanobis distance under unreliable Kalman Filter predictions. Moreover, the association strategy is enhanced through a comprehensive cost matrix that adaptively fuses spatial overlap, appearance descriptors, velocity direction, and shape consistency. Furthermore, a Tentative Track Recovery (TTR) strategy combined with a post-processing interpolation strategy is proposed to mitigate information loss and maintain trajectory continuity from the very inception of identity initialization. Extensive experiments demonstrate that the integrated pipeline reaches state-of-the-art performance on MOT17 and achieves highly competitive results within the leading group of Tracking-by-Detection (TBD) methods on both MOT20 and DanceTrack benchmarks. Code and models are available at https://github.com/HuynhNghiaKHMT/PrismTrack.