Oriented Road Marking Detection in MLS Point Cloud Intensity Images Using Channel and Spatial Attention

Accurate detection of road markings is critically important in fields such as autonomous driving technology, high-precision mapping, and intelligent transportation systems. Unlike traditional object detection tasks, road marking detection faces many challenges, including significant direction changes, complex backgrounds, and diverse types of road markings. To address these issues, this article proposes a deep learning approach based on mobile laser scanning point cloud intensity images to accurately detect road markings with arbitrary orientations. First, we employ an oriented bounding box to precisely represent the direction and location of road markings, thereby improving detection accuracy. Additionally, the channel semantic enhanced module (SEM) is designed to enhance the feature representation capacity of the backbone network, effectively reducing the impact of complex backgrounds on foreground targets. Finally, the spatial SEM is introduced to enhance the feature pyramid network's ability to capture contextual information. The module consists of the large selective kernel network (LSKNet) and the SEM. LSKNet extracts contextual information around the target by adaptively selecting receptive fields of various scales, while SEM, a lightweight semantic segmentation branch, further strengthens the feature representation ability of this information by incorporating a semantic supervision mechanism. The proposed method achieves an mAP of 84.7% on Test Set I, representing a 3.2% improvement over the baseline model, while maintaining high efficiency, offering a robust and reliable solution for real-time road marking detection tasks.