On Thursday, 30th of July 2020, 09:00 AM, M.Sc. student “Dunia Ali Abdel-Hamza” defended her thesis entitled:
“Prediction of Traffic Flow Based on Deep Learning Approach”
Via Zoom meetings Online platform.
The discussion committee included:
Prof.Dr. Mahmoud Khalil Ibrahim / College of Information Engineering. /Chairman.
Assistant Prof.Dr.Ammar Abdel-Malik Abdel-Karim / College of Information Engineering / Member.
Assistant Prof.Dr. Muhammad Issam Younis / University of Baghdad / College of Engineering/member.
Assistant Prof.Dr. Ammar Dawood Jasim / College of Information Engineering. Member and supervisor.
Abstract :
License plate Recognition (LPR) is used in many applications, such as estimation of parking traffic, border control, and tolling on the motorway. Prediction accuracy and LPR speed are important. Recent advances in deep learning (DL) have improved its ability to solve complex visual recognition tasks. Used DL to improve accuracy and speed with which to solve the LPR problem. The goal of this thesis was to propose a method which used DL techniques to predict traffic flow by solving the problem of LPR. A system that used convolutional neural networks (CNNs) for object detection was proposed to solve the LPR challenge. The first network was retrained to identify license plates, and the second, using the speeded-up robust feature (SURF) approach to recognize numbers inside the license plates the first network detected. The approach proposed was new which has never been tested at the LPR project before. The assessment of the proposed method resulted in overall predictability of 96 %. In training, the license plate detection achieved 99.2 % accuracy, and 96 % SURF accuracy for character recognition.
When the mechanism proposed was running on a GPU. This for deep learning, in machine learning (ML) also, has been used K nearest neighbors (KNN) a method for LPR, The KNN algorithm used for detection and recognition LPR with accuracy 85% in training and 90% in testing, to be compared between ML and DL. The proposed system achieved very high accuracy prediction and outperformed, all other methods considering the processing speed when using a graphics processing unit (GPU). This indicates that using CNNs for deep object detection is a good solution to the LPR problem.
At the end of the session, the thesis was accepted with minor corrections within a period of one month.