Context-aware and boundary-optimized model for road marking instance segmentation using MLS point cloud intensity images

Accurate road marking extraction is essential for advancing digital transportation systems, autonomous vehicles, and high-definition maps. Although existing methods focus on extracting high-precision road markings from Mobile Laser Scanning (MLS) point clouds, they still face challenges in practical applications, including indistinguishable instances, category ambiguity, and incomplete boundary segmentation, which collectively limit their overall performance. To address these challenges, we convert MLS point clouds into intensity images and propose a deep learning model that integrates context modelling and boundary refinement for accurate instance segmentation of road markings. First, a feature enhancement module (FEM) is designed to improve road marking representation by learning dependencies across channel and spatial dimensions. Second, a multiscale context-aware module (MSCAM) is constructed to enhance the model's capacity to identify diverse marking types by aggregating semantic information from multi-scale distant regions. Lastly, a PointRend module (PRM) is introduced to adaptively select key points for prediction, generating high-quality boundary masks. Experiments on a newly constructed dataset reveal that, compared with state-of-the-art instance segmentation models, our method offers substantial performance advantages. The model accurately detects and segments eight categories of road markings, achieving 77.1% APb and 59.8% APm on Test Set 1.