Real-Time Bridge Weigh-in-Motion with Computer Vision and Structural Response Under Variable Speed and Mixed Traffic

A field-oriented, real-time implementation of the previously developed vision-based bridge weigh-in-motion (V-BWIM) framework is presented for vehicle-load monitoring under variable-speed and mixed-traffic conditions. Vehicle and wheel positions are obtained through YOLOv5-based object detection, binocular measurement, and coordinate transformation, whereas bridge responses are measured by strain gauges and synchronized with the vision-derived axle trajectories. A field-image dataset constructed from full-scale bridge experiments is used to compare object-detection performance, inference efficiency, and model complexity, and YOLOv5s is selected for field implementation. The calibrated bridge influence line is then combined with synchronized axle positions and structural responses for axle-load identification. Controlled field tests on a simply supported bridge quantitatively evaluate vehicle positioning and load identification under constant-speed, variable-speed, and two-vehicle car-following conditions. A further continuous-beam bridge experiment demonstrates vehicle-information estimation under random traffic. Because independent reference weights were unavailable for the randomly passing vehicles, the continuous-bridge results are interpreted as a field demonstration rather than an independent validation of load-identification accuracy.