Real-Time Scale-Invariant License Plate Detection Using Cascade Classifiers

摘要

This paper presents an online scale-invariant license plate detection (LPD) system with high accuracy for the automatic license plate recognition (ALPR) systems. A dataset of Persian plates is accumulated with more than 44,000 images of plates and 9000 frames of real world roads. For the plate detection and localization, a multi-stage classifier is trained with local binary pattern (LBP) features and a multi-scale algorithm to detect plates with any size within a frame. Besides, we proposed multiple algorithms to boost the performance and accuracy of our solution, including two-stage detection, background subtraction for non-moving areas elimination, and a sophisticated method for estimating the size of the plates in each part of the frames based on linear regression. In our most precise method we achieved 99.6% of accuracy, with detection rate of 83.68 fps. We also proposed several combinations of our algorithms to speeding up the process to 100 fps.