A deep learning-based approach for axle counter in free-flow tolling systems

Enhancements in the structural and operational aspects of transportation are important for achieving high-quality mobility. Toll plazas are commonly known as a potential bottleneck stretch, as they tend to interfere with the normality of the flow due to the charging points. Focusing on the automation of toll plazas, this research presents the development of an axle counter to compose a free-flow toll collection system. The axle counter is responsible for the interpretation of images through algorithms based on computer vision to determine the number of axles of vehicles crossing in front of a camera. The You Only Look Once (YOLO) model was employed in the first step to identify vehicle wheels. Considering that several versions of this model are available, to select the best model, YOLOv5, YOLOv6, YOLOv7, and YOLOv8 were compared. The YOLOv5m achieved the best result with precision and recall of 99.40% and 98.20%, respectively. A passage manager was developed thereafter to verify when a vehicle passes in front of the camera and store the corresponding frames. These frames are then used by the image reconstruction module which creates an image of the complete vehicle containing all axles. From the sequence of frames, the proposed method is able to identify when a vehicle was passing through the scene, count the number of axles, and automatically generate the appropriate charge to be applied to the vehicle.