Date of Award

2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Dr. Kaushik Roy

Abstract

Traffic Sign Recognition System (TSRS) is an Advanced Driver Assistance System (ADAS) that helps drivers with perception to ensure road safety. Two main activities are performed in TSRS: detection and classification. The detection aspect involves localizing traffic signs in an image frame while the classification aspect deals with recognizing the class of the detected sign. Research in this area has mainly focused on German, Belgium, Sweden, Chinese, and several other datasets using different approaches. However, limited research has been conducted using U.S. traffic signs; the ones that have been conducted are mostly concerned with speed limit signs recognition. This work expands the classification of U.S. traffic signs to cover all the publicly available classes. Convolutional Neural Networks (CNN) have shown a lot of success on European datasets. One key issue with CNN is that it requires a lot of data for training. This research introduces a new model, called NuNet, with a new dataset, CIB TS V1. The model is used to classify the CIB dataset and LISA benchmark (Møgelmose et al., 2012). Results from running NuNet on LISA and CIB are then compared with those of a modified VGGNet. The new model trains and converges faster than VGGNet and is adaptable to both large and sparse datasets. Experiments conducted with the VGGNet show training and validation accuracies of 99.93% and 99.83, respectively on the LISA dataset. However, it overfits on the CIB dataset with training and validation accuracies of 100% and 96.92%, respectively. This is because the deep net cannot generalize well on small datasets and thereby learns noise. NuNet, on the other hand, generalizes well on smaller datasets recording accuracies of 99.73% and 99.83% on LISA for the training and validation sets respectively, and 100% for both training and validation sets on the CIB dataset. NuNet trains for 4 hours on LISA and an hour on CIB whereas VGGNet trains for 23 hours on LISA and 8 hours on CIB.

Download

Share

COinS