BiCAT: bidirectional cross-attention for asynchronous heterogeneous multi-source trajectory association

Le résumé fourni par la source

In recent years, multi-source, multi-target trajectory association is critical for air traffic management and intelligent transportation systems. Yet prevailing methods rely on rigid temporal alignment and hand-crafted rules; they deteriorate under cross-modal asynchrony, heterogeneous measurements, and high-density scenes. This paper presents BiCAT (Bidirectional Cross-Attention for Asynchronous Heterogeneous Multi-Source Trajectory Association), an end-to-end framework with bidirectional cross-attention for asynchronous, heterogeneous, multi-source association between active radar and infrared sensors. BiCAT dispenses with interpolation and dimensional pre-alignment by dynamically aligning in feature space. A dual-path, mask-aware attention module dynamically suppresses invalid or missing observations. A discriminative fusion classifier further enhances association robustness by jointly exploiting both similarity and divergence between sensor modalities. In crossing-trajectory scenarios, BiCAT attains 94% precision, outperforming SGFA by 9.3%; in parallel-queue scenarios, it achieves 92.6% recall, surpassing GCD-TA by 19.8%. These results demonstrate consistent gains in complex, high-density environments and provide reliable support for multi-source fusion.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.