Speaker
Description
Distributed Acoustic Sensing (DAS) enables continuous monitoring of traffic activity over long distances. However, noise contamination and the lack of labeled data limit the scalability of existing approaches. This study proposes a framework that integrates a physics-aware learnable filtering module with a data-driven refinement strategy for vehicle detection in DAS signals. The filtered outputs are used to generate and progressively improve proxy labels, enabling robust segmentation without manual annotations. The extracted vehicle regions are further utilized for speed estimation. Results on real-world data demonstrate stable vehicle detection under noisy conditions and reliable speed estimation performance. The findings highlight the potential of combining physics-informed filtering with adaptive learning for scalable DAS-based traffic monitoring.