Automated Traffic Signal Recognition in Distributed Acoustic Sensing Data via Deep Learning

Abstract Distributed Acoustic Sensing (DAS), an advanced vibration-sensing technology, shows immense promise for data-centric urban surveillance, notably in tracking vehicle speeds. The extraction of vehicle information from DAS data demands real-time processing. Nonetheless, current techniques face challenges in the precise and automated interpretation of vehicle signals. This study presents an integrated two-phase deep learning framework meticulously designed to facilitate the real-time and automated detection of vehicle speeds and directions using Distributed Acoustic Sensing (DAS) data. Initially, the approach employs a contrastive learning-based model to process DAS signals, eliminating disturbances within the profile. The YOLOv8 detection model subsequently accurately detects the processed DAS signals. The vehicle’s speed and direction are ultimately ascertained by harnessing the positional data derived from the bounding box’s coordinates. Within real-world test scenarios, the method introduced can precisely identify vehicles originating from various directions and sources. The strategy demonstrates robust generalizability even in complex scenarios characterized by intense interference and the presence of multiple vehicles. From a quantitative assessment perspective, the system processes DAS data for 60-second intervals in an average of 4.89 seconds, achieving an accuracy rate of 90.07%, which satisfies the demands for real-time vehicle signal detection. The approach presented in this article offers crucial guidance for the real-time, automatic detection of vehicle signals in DAS systems.