Vehicle Trajectory Extraction From Weak and Discontinuous Quasi-Static DAS Signals Using a Multiscale Context Pix2pix

In recent years, optical fiber distributed acoustic sensing (DAS) technology has gained prominence in vehicle detection due to its low deployment cost, high spatial resolution, and ultra-long-distance measurement. However, various adverse factors, including strong environmental noise, DAS fading noise, lightweight vehicles, low speed, and poor coupling between the fiber cable and the ground, cause vehicle signals to weaken and become discontinuous. These issues limit the measurement sensitivity and reliability of DAS systems, making it challenging to accurately extract vehicle trajectories using existing algorithms. This paper introduces a Multi-Scale Context Pix2pix (MCPix2pix) generative network model embedded into the DAS signal measurement pipeline to enhance the system’s ability to detect and reconstruct weak and discontinuous quasi-static vehicle signals. Adversarial learning between the generator and discriminator, effectively repairs fragmented vehicle trajectories. Furthermore, multi-scale dilated convolutions and corresponding loss functions expand the receptive field and improve the capture of vehicle features across multiple spatial and temporal scales. Field experiments conducted with vehicles on campus demonstrated that applying the MCPix2pix model to DAS quasi-static signals can reliably extract discontinuous trajectories of low-speed, lightweight vehicles, thereby extending the functional sensitivity range of DAS and improving its effectiveness as a distributed sensing instrument in real-world applications.