Use this URL to cite or link to this record in EThOS:
Title: Detection and recognition of traffic scene objects with deep learning
Author: Qian, Rongqiang
ISNI:       0000 0004 7428 5758
Awarding Body: University of Liverpool
Current Institution: University of Liverpool
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Mobility is an element that is highly related to the development of society and the quality of individual life. Through mass of automobile production and traffic infrastructure construction, advanced countries have reached a high degree of individual mobility. In order to increase the efficiency, convenience and safety of mobility, advanced traffic infrastructure construction, transportation systems and automobiles should be developed. Among all the systems for modern automobiles, cameras based assistance systems are one of the most important components. Recently, with the development of driver assistance systems and autonomous cars, detection and recognition of traffic scene objects based on computer vision become more and more indispensable. On the other hand, the deep learning methods, in particular convolutional neural networks have achieved excellent performance in a variety of computer vision tasks. This thesis mainly presents the contributions to the computer vision and deep learning methods for traffic scene objects detection and recognition. The first approach develops numbers of methods for traffic sign detection and recognition. For traffic sign detection, template matching is applied with new features extended from chain code. Moreover, the region based convolutional neural networks are applied for detecting traffic signs painted on road surface. For traffic sign recognition, convolutional neural networks with a variety of architectures are trained with different training algorithms. The second approach focuses on the detection related to traffic text. A novel license plate detection framework is developed that is able to improve detection performance by simultaneously completing detection and segmentation. Due to the larger number and complex layout of Chinese characters, Chinese traffic text detection faces more challenges than English text detection. Therefore, Chinese traffic texts are detected by applying convolutional neural networks and directed acyclic graph. The final approach develops a method for pedestrian attribute classification. Generally, there are irrelevant elements included in features of convolutional neural networks. In order to improve classification performance, a novel feature selection algorithm is developed to refine features of convolutional neural networks.
Supervisor: Yue, yong ; Zhang, Bailing ; Coenen, Frans Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral