Robust DAS vehicle trajectory extraction based on spatial cross-section semantic segmentation

Distributed Acoustic Sensing (DAS) technology holds significant promise for traffic monitoring, yet extracting high-precision vehicle trajectories remains a major challenge due to low signal-to-noise ratios and prevalent linear noise artifacts. Traditional methods relying on 2D spatio-temporal morphology, such as Radon or Hough transforms, are often susceptible to these artifacts and struggle with weak signals. To address these limitations, this paper proposes D-TrajNet, an innovative end-to-end trajectory extraction pipeline. The core of this method employs a novel 1D semantic segmentation model (DAS-1D-SegNet) combining 1D-CNN and Bi-LSTM to learn specific vehicle ”waveform fingerprints” directly from spatial cross-sections, effectively distinguishing signals from noise without relying on 2D geometric shapes. Furthermore, a geometric post-processing workflow integrating DBSCAN and RANSAC is introduced to achieve adaptive denoising, multi-vehicle separation, and the automatic repair of trajectory discontinuities caused by signal interruptions. Experimental results demonstrate that D-TrajNet significantly outperforms traditional signal processing baselines, improving the detection F1-Score by up to 32.7% and reducing positioning errors. Qualitative comparisons confirm the system’s superior robustness against linear artifacts and its effectiveness in reconstructing smooth, continuous trajectories for Intelligent Transportation Systems.