孫車牌: 基於深度學習之車牌辨識系統

Abstract

本論文提出一個專用於機車上的車牌辨識(License Plate Recognition)系統解決方案。搭載工業電腦以及人工智慧(AI: Artificial Intelligence)加速卡，可在載具行進期間，同時進行實時的車牌辨識。本篇論文透過自行改良的深度學習網路架構，提高機車在行駛途中攝影機所拍下的影像辨識準確率。此外，本研究更提出整個系統安裝至機車時的配置以及設計參考。為了更好的使用者體驗，本篇研究將車牌辨識程式整合成一系列應用程式介面(API: Application Programming Interface)，實際開發出一套完整的「AI人工智能警用機車影像識別偵查系統」。配合應用程式介面的開發，此系統可以實時處理兩部攝影機的影像輸入，同時對前方車道以及右側路邊停車區域進行車牌辨識。

We propose SunLPR, a license plate recognition system solution dedicated to common scooter. Equipped with IPC (Industrial Personal Computer) and AI (Artificial Intelligence) acceleration card (Mustang V-100 Movidius Mx8). Our SunLPR can perform real-time license plate inference tasks when scooter is moving. A modified deep learning network also implicitly improves the accuracy of plate number recognition. And a brief design of the AI scooter was introduced, including cameras and battery. In this thesis, SunLPR was merged into an API (Application Programming Interface) for a better user experience. This AI system can deal with two camera’s video input in real-time and tracking the license plates in front and on the right.