Automated Toll Plaza

Abstract

Operations at toll plazas frequently face difficulties like inefficiency, mistakes made when allocating tolls manually, and the requirement for extra infrastructure. Current solutions, such as M-TAG require substantial alterations to both toll booths and vehicles. Our project uses the already underutilized CCTV infrastructure to automate toll collection to address these problems. For precise identification of vehicle and number plate types, we incorporate an ANPR (Automatic Number Plate Recognition)-based OCR (Optical Character Recognition) pipeline using YOLO v8 models. This method improves dependability and efficiency without requiring new infrastructure. By pre-processing the extracted data, paddle OCR ensures high accuracy even in the face of obstacles like handwritten, occluded, or broken license plates. To ensure robustness, our solution is tested using a Pakistani dataset with a range of weather conditions. It works well on open-loop roads such as GT Road and in situations where M-TAG is not present. Our system can also be used for surveillance and traffic management because it has a web application for data management and real-time monitoring. This creative solution meets the unique requirements of contemporary transportation systems while reducing the operation of the toll plaza. We collected unique data from local toll plazas on the M-1 motorway in Islamabad, such as Tarnol and Sarang Jani. The data set consists of 2,000 images from five types of vehicles. Each image is manually annotated for vehicle type, license plate, and characters/digits. We evaluated YOLOv8 for each of the three tasks.Tiny YOLOv3 and Tiny YOLOv4 are lightweight versions of YOLO that are evaluated for real-time implementation on Raspberry Pi. YOLOv4 achieves the highest Mean Average Precision (mAP@0.5) of 98.8 percent for vehicle type recognition, 98.5 percent for plate recognition, and 98% for plate reading. Tiny YOLOv4 has a lower mAP of 97.1 percent, 97.4 percent, and 93.7 percent, respectively.