Siddharth Agrawal Wins Best Conference Paper Award for his Study on Identification of Truck Licence Plates for Weighbridge Automation

Siddharth Agrawal, BS (Hons) in Computer Science, School of Arts and Sciences, has won the Best Conference Paper Award at the 28th IEEE International Conference on Mechatronics and Machine Vision in Practice (M2VIP) 2022. He presented his paper titled Indian Commercial Truck License Plate Detection and Recognition for Weighbridge Automation at Northeast University, Boston, USA. Keyur D Joshi, Assistant Professor, School of Engineering and Applied Science, is the co-author on the paper and Siddharth’s mentor.

Automating commercial vehicle licence plate recognition is essential for logistics management, weighbridge automation and other intelligent transportation system tasks. Unfortunately, Indian licence plate datasets are small and inadequate. Siddharth says, “Our project was to develop a database and effective models to recognise and detect Indian commercial truck licence plates. Currently, this is a serious challenge as we have handwritten plates, various font styles, noise, visual obstructions, distorted and contorted plates, and illumination issues. We used state-of-the-art models like real-time Object Detection (YOLOv7) and Scene Text Recognition (PARSeq).” Through modifications made to the recognition model, such as increasing the resolution and aspect ratio of the input, developing novel annotation techniques to target Indian data, and generating synthetic datasets, Siddharth and Professor Joshi successfully increased the accuracy rate of recognition from 58.96 per cent to 95.82 per cent, with an F-1 score of 99 per cent for detecting the licence plates.

Professor Joshi says, “The task was difficult mainly due to the hand-painted licence plates with different styles and noises. A single incorrect letter can lead to the entire plate being recognized incorrectly, and reduce the accuracy. The low light and obstructions on licence plates contributed more towards the difficulty level of the task. The findings of this study will go a long way in assisting the process of weighbridge automation, a fully automated service with minimal to no human intervention.”