A fast and environmentally adaptable method for identifying heavy vehicles weight distribution through video and Electronic Toll Collection data fusion

An important engineering requirement (especially for heavy vehicles) is the identification of vehicle weight distribution. However, (1) most bridges are located in complex outdoor environments that require a high degree of environmental adaptability when identifying. Furthermore, (2) real-time video input generates a large amount of computation. This can cause significant delays in the identification process. Therefore, a fast and environmentally adaptive heavy vehicle weight distribution identification method is proposed in this paper. In this method, (1) a triggering device and a light object detection network (vehicle detection and number plate detection) are first designed to improve the identification efficiency, then (2) the training set of the object detection network is expanded by images conversion methods based on generative adversarial network (GAN) to satisfy environmental diversity and label equilibrium, and finally (3) the heavy vehicle distribution is obtained by vehicle tracking based on vehicle feature matching and Electronic Toll Collection vehicle weight association based on number plate matching. The proposed method is successfully applied to a single-tower cable-stayed bridge. This precision and recall of vehicle and number plate recognition are both above 95.6%. In addition, the designed lightweight network YOLOv5s-p improves the prediction speed by about 30% compared to YOLOv5s, and the trigger device reduces the computation tasks by about 70%.