| dc.description.abstract | Urban roads, as one of the essential transportation infrastructures, provide considerable motivations for rapid urban sprawl and bring notable economic and social benefits. Accurate and efficient extraction of road information plays a significant role in the development of autonomous vehicles (AVs) and high-definition (HD) maps. Mobile laser scanning (MLS) systems have been widely used for many transportation-related studies and applications in road inventory, including road object detection, pavement inspection, road marking segmentation and classification, and road boundary extraction, benefiting from their large-scale data coverage, high surveying flexibility, high measurement accuracy, and reduced weather sensitivity. Road information from MLS point clouds is significant for road infrastructure planning and maintenance, and have an important impact on transportation-related policymaking, driving behaviour regulation, and traffic efficiency enhancement.
Compared to the existing threshold-based and rule-based road information extraction methods, deep learning methods have demonstrated superior performance in 3D road object segmentation and classification tasks. However, three main challenges remain that impede deep learning methods for precisely and robustly extracting road information from MLS point clouds. (1) Point clouds obtained from MLS systems are always in large-volume and irregular formats, which has presented significant challenges for managing and processing such massive unstructured points. (2) Variations in point density and intensity are inevitable because of the profiling scanning mechanism of MLS systems. (3) Due to occlusions and the limited scanning range of onboard sensors, some road objects are incomplete, which considerably degrades the performance of threshold-based methods to extract road information.
To deal with these challenges, this doctoral thesis proposes several deep neural networks that encode inherent point cloud features and extract road information. These novel deep learning models have been tested by several datasets to deliver robust and accurate road information extraction results compared to state-of-the-art deep learning methods in complex urban environments. First, an end-to-end feature extraction framework for 3D point cloud segmentation is proposed using dynamic point-wise convolutional operations at multiple scales. This framework is less sensitive to data distribution and computational power. Second, a capsule-based deep learning framework to extract and classify road markings is developed to update road information and support HD maps. It demonstrates the practical application of combining capsule networks with hierarchical feature encodings of georeferenced feature images. Third, a novel deep learning framework for road boundary completion is developed using MLS point clouds and satellite imagery, based on the U-shaped network and the conditional deep convolutional generative adversarial network (c-DCGAN). Empirical evidence obtained from experiments compared with state-of-the-art methods demonstrates the superior performance of the proposed models in road object semantic segmentation, road marking extraction and classification, and road boundary completion tasks. | en |
