End-to-End License Plate Recognition System with Optimized Inference on Mobile Security Robots

Automatic license plate recognition (ALPR) systems are increasingly used to solve issues related to surveillance and security. However, these systems assume constrained recognition scenarios, thereby restricting their practical use. In this paper, we address the challenge of recognizing vehicle license plates (LPs) using a mobile security robot. We present a novel license plate recognition (LPR) model, called vision transformer based LP recognizer (ViTLPR), that benefits from recent advances in deep learning. ViTLPR is based on self-attention mechanism to read character sequences on LPs. To ease the ViTLPR deployment and improve its inference speed, we propose an optimization process using TensorRT. Combined with on-the-shelf YOLOv7x model, we build a two-stage ALPR system and deploy it on a commercial security robot. As an additional contribution, we provide a LP dataset, called PGTLP-v2, collected from surveillance robot patrolling several plants. The PGTLP-v2 dataset has multiple features to cover the in the wild scenario. Experimental results show that the system is competitive compared with state-of-the-art methods.