SynTrackThinking improves multimodal multi-object tracking for autonomous driving through frequency-aware fusion and temporal contrastive learning

作者：Zhihao Su, Wei Chang · 发表于：Scientific Reports · 年份：2026 · DOI：10.1038/s41598-026-44182-4 · 研究领域：Video Surveillance and Tracking Methods、Gaze Tracking and Assistive Technology、Autonomous Vehicle Technology and Safety

Multimodal multi-object tracking now constitutes a foundational capability of autonomous driving platforms, allowing intelligent vehicles to perceive and continuously track dynamic agents through the joint exploitation of visual, auditory, and linguistic signals. Nevertheless, the problem remains hindered by two enduring challenges. One lies in the limited effectiveness of multimodal fusion-most notably the neglect of frequency-domain cues when representation learning is confined to spatial processing alone. The other concerns the instability of temporal semantic alignment, a weakness that often manifests as cumulative tracking drift over consecutive frames. To address these issues, prior studies have explored spatial-frequency fusion mechanisms alongside contrastive learning strategies designed to align heterogeneous modality embeddings. Although such efforts have yielded measurable improvements, their capacity to accommodate complex cross-modal interactions and volatile long-term temporal dependencies remains constrained. Motivated by these limitations, we propose SynTrackThinking, an integrated multimodal tracking framework that tightly couples frequency-aware representation learning with contrastive temporal reasoning. The framework is built upon two core components: an explainable multimodal-domain cross-attention fusion module (EMCFM), which enhances representational expressiveness via coordinated attention across spatial and frequency domains, and a multimodal contrast...