Licence Plate Identification and Recognition for Non-Helmeted Motorcyclists using Light-weight Convolution Neural Network

Abstract

Motorcycle accidents have been rapidly increasing in many countries. The helmet is the main safety equipment of motorcyclists, but many drivers do not use it. Helmets are essential for the safety of a motorcycle rider. Hence, detecting and extracting licence plate of the motorcycle in which riders have not wear helmet becomes a crucial task. Many methods have been proposed to detect and extract the licence plate; however, due to poor video quality and non-uniform illumination, licence plate detection becomes a difficult task. Recently, due to the advancement in graphical processing units (GPUs) and larger datasets, deep learning based models have obtained remarkable performance in the object detection task. One such model is single shot detection (SSD) which classify and detect real-time objects precisely. In this paper, we propose an end-to-end approach for detecting and extracting a licence plate of the motorcycle. Here, we use a MobileNet based SSD model to detect License plates as MobileNet i.e., a light-weight CNN model which is more suitable for mobile and embedded vision applications to obtain fast operation. We also prepare a dataset of Indian motorcycle licence plates which consists of 1524 images to train and validate the SSD model. From experiments, we found that the detection module detects the Indian motorcycle licence plate accurately. Once the License plates are detected, the detected licence plate is extracted and the characters of the extracted licence plate are recognized through optical character recognition (OCR) module.