Fiber-optic distributed acoustic sensing (DAS) offers a promising solution for continuous traffic monitoring; however, its widespread deployment is often hindered by poor signal quality, resulting in fragmented and faint vehicle trajectories. Existing techniques − including conventional signal processing and deep learning models − struggle to accurately reconstruct trajectories and estimate traffic parameters under such challenging conditions. To overcome these limitations, we propose the DAS-hierarchical vehicle estimation network (DAS-HiVENet), an end-to-end framework that fundamentally advances the state-of-the-art through three key innovations: a two-stage preprocessing pipeline for noise suppression and trajectory preservation; a novel generative adversarial network (GAN) with an enhanced U-shaped convolutional neural network (U-net) generator to reconstruct high-fidelity trajectories from degraded inputs; a rotated-you only look once (R-YOLO) detector using oriented bounding boxes to accurately detect slanted trajectories. Extensive field evaluations on multiple expressways confirm that it surpasses existing methods with breakthrough performance: a trajectory intersection over union (IoU) of 0.7076, vehicle counting detection rate of 96.7%, and speed estimation errors as low as 1.422 km/h for the mean absolute error (MAE) and 1.796% for the mean absolute percentage error (MAPE) over 30 minutes. Even in challenging bridge scenarios with severe trajectory adhesion, DAS-HiVENet maintains an over 96% detection rate and under 4% MAPE in speed estimation − significantly outperforming alternatives.
Wu et al. (2026) studied this question.