The objective of this study is to advance automated pavement crack segmentation by addressing two major bottlenecks that hinder the performance of convolution neural networks (CNNs) on this task, which are the robustness against diverse scenarios of real-world pavement data and the poor detail-preserving capability of the CNNs. To achieve this objective, we propose two methodologies. First, we propose to improve the robustness by leveraging the local geometry inherent in pavement data. To this end, we propose the Dense Local Geometry (DLG) features. These features are incorporated as an additional input to the CNN, providing the network with a sense of local geometry that is tailored specifically to cracks during training. Second, we propose to enhance the detail-preserving capability by introducing the Boundary Enhancement (BE) loss. The BE loss constructs a supervisory signal on the boundaries of segmentation regions, enhancing the network’s capability to perceive small deviations and disconnectivity at the boundary of cracks during training. The outcomes of this research will support transportation agencies in achieving detailed pavement condition assessments by providing high-granularity and precise crack segmentation from pavement range images collected with the emerging 3D laser technology.