Vehicle identification using optimised ALPR.

Abstract

Vehicles are a common sight on the road. Tracking and monitoring suspicious vehicles for identification due to high similarity in structure and form leads to difficulties in differentiating between them. The unique identity of a vehicle, the license plate is used here for this purpose. License plate detection is considered as an object detection task. Transfer learning on pre-trained state of art object detection models is an approach, which can perform this with better accuracy in terms of mean average precision. However, setting the right hyper-parameters needs multiple experiments. In this research, an evolutionary algorithm, genetic algorithm is used, which can optimize the hyper-parameters to achieve the best accuracy for the object detection model, YOLOv5. Further, the license plate was identified using OCR. This study concluded that hyper-parameter tuning achieved high accuracy in terms of mean average precision, achieving 98.25%, compared to 80% in initial parameter set providing an automated optimization. This license plate detected can be stored in a secure location and retrieved for re-identification. A decentralized storage or a secure cloud can be used to store the license plate. The application of this is most relevant to surveillance in high security locations where suspicious vehicles must be tracked.