DEVELOPMENT OF AN AUTOMATED TRAFFIC MONITORING SYSTEM WITH LICENSE PLATE RECOGNITION AND TALLYING

Abstract

Automated Traffic Monitoring is integral to the development of a functional modern Intelligent Transportation System (ITS). Traffic data such as traffic flow rate, passage time, vehicles, vehicle count, and frequency of a particular vehicle type obtained from these monitoring processes are required by transport modelling and urban planning researchers, law enforcement, and road maintenance agencies but are not available as traffic monitoring systems are not locally available. This research, therefore, develops an Automated Traffic Monitoring System with license plate recognition and electronic Tallying capabilities. The developed system is made up of six (6) units viz: Camera modules, Raspberry Pi-based system unit, Radio Frequency Identification (RFID) Card reader modules, Personal Integrated Circuit Cards (PICC) and Key-fobs, ATMEGA328P based embedded system unit, and power supply unit, and implements computer vision and IoT based traffic monitoring techniques. The system’s operational algorithm was realized using the python and C++ programming languages on the Raspberry Pi and the embedded system respectively. A validation dataset of 120 images spread across four different vehicle classes (car, SUV, bus, and truck) was used to validate the system’s deep learning object detection pipeline, the tesseract-based license plate optical character recognition (OCR) engine, and the vehicle classifier. The object detector pipeline achieved an accuracy of 99.16% on detecting vehicle instances in the dataset and an accuracy of 100% on the detection of plate instances in the dataset. The OCR engine recognized all plate number characters correctly for 78 license plates in the dataset, recognized one character wrongly for 27 license plates in the dataset, and recognized two characters and more wrongly for 16 license plates in the dataset. The vehicle classifier achieved an accuracy of 87% on vehicle class prediction on all images in the dataset. The developed system was tested at a traffic gate where it achieved a near real time average frame rate of 2.5 frames per second. It achieved 100% accuracy on the detection of vehicle instance at distances less than 10 meters from the system, 75% accuracy on the detection of license plates at distances less than 5 meters from the system, 77% accuracy on recognizing all detected license plate character, and 100% accuracy on the prediction of vehicle classes for the vehicle observed at the traffic point. The developed system successfully logged all traffic information collated at the traffic gate to a database and can therefore serve as a scalable source of localized traffic information.