- Title
- Road features extraction based on fusion of point cloud and camera image detection
- Creator
- Wanady, Irene
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2024
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Road safety is closely related to road information inventory as it provides essential data for understanding and enhancing safety conditions on roads. Through cataloguing of up-to-date road features such as lane markings, signs, and roadside objects, road inventory enables transportation authorities to identify potential hazards and plan targeted safety interventions. Furthermore, it allows analysis of past road data, which can assist transportation officials to detect accident-prone areas and implement strategies to reduce risks and improve overall safety. Ultimately, road information inventory plays a crucial role in supporting proactive road safety initiatives and minimising accidents and injuries on roads. Road information inventory involves identifying various objects on road, therefore road feature extraction plays a crucial role in the road inventory process by allowing efficient and accurate detection as well as classification of road infrastructure elements. With the rapid development of technology, mobile laser scanning (MLS) systems have become a widely used tool in many transportation-related applications including the road inventory process. MLS systems allow efficient and accurate capture of road scenes in the form of 3D point cloud data and 2D camera imagery. Many researchers have focused on developing feature extraction algorithms based on either point cloud or image data. However, the information stored in both these sets of data have their own limitations which restricts the capability of the algorithms designed when exploiting either one of the data types. Therefore, the research conducted in this thesis will be focused on developing road feature extraction algorithms which leverage the information from both 3D point cloud and 2D camera imagery. Different approaches are explored, including conventional image processing methods, computer vision techniques, and deep learning networks, to show the advantages when combining both data types to detect road features. In this thesis, several pre-processing algorithms are first presented, including a ground extraction algorithm which is often an essential component in most road feature extraction methods. Next, a two-stage fusion framework to detect roadside objects, such as poles and walls, using mostly traditional techniques is developed. The use of a deep learning network is then explored with the introduction of a hybrid approach. The proposed methodology combines conventional image processing methods with deep learning architectures to detect road symbols/markings by utilising data from both 3D point clouds and 2D camera images. Finally, with the increased interest in object detection using deep learning network, a novel end-to-end network architecture which performs segmentation on small and challenging road objects is proposed. It is demonstrated that the proposed fusion frameworks/methods can accurately, and effectively, perform road feature extraction process to assist with the challenge of creating a comprehensive road information inventory.
- Subject
- road features extraction; road safety; camera image detection; point cloud
- Identifier
- http://hdl.handle.net/1959.13/1512450
- Identifier
- uon:56618
- Rights
- This thesis is currently under embargo and will be available from 25.09.2025. Copyright 2024 Irene Wanady
- Language
- eng
- Hits: 45
- Visitors: 40
- Downloads: 0
| Thumbnail | File | Description | Size | Format |
|---|