A machine learning-based noise pollution detection system for urban traffic

Abstract Urban road-traffic noise is commonly reported using long-term averaged indicators, which limits analysis of short-duration high-level events and their association with individual vehicles. We present an event-driven acoustic–vision measurement framework that converts short-window A-weighted equivalent-level exceedances (L_Aeq,0.1 s) into time-stamped, vehicle-attributed records. Audio from two directional microphones is sampled at 48 kHz and processed in 100 ms windows. A digital A-weighting filter is implemented in software following the IEC 61672-1:2013A-weighting specification to estimate L_Aeq,0.1 s at the receiver. When a user-defined threshold is exceeded, the system triggers synchronised image capture and performs licence-plate detection (opting YOLOv5) and optical character recognition (OCR) to document the corresponding vehicle. A multi-threaded implementation processes audio and video concurrently on a laptop, while the resulting database links exceedance time, lane/channel metadata, geolocation and plate-text outputs to support operationally actionable vehicle-level attribution. The detector achieved mAP@0.5 = 0.988 (mAP@0.5:0.95 = 0.73). In peak-hour field deployment on a two-lane urban road, plate-detection success reached 94.2%, and OCR accuracy was 82.6% for exceedance frames.