Abstract: Objectives： Efficient and accurate road marking extraction is fundamental for vehicle localization, environmental perception, and high-definition map construction. However, vehicle- borne images are usually large in volume and costly to process, making it difficult for existing methods to achieve a favorable balance between segmentation accuracy and computational efficiency. To address this challenge, this paper proposes a lightweight road markings network (LRMSNet) to accurately extract road markings from vehicle-borne images with low parameter complexity and computational cost. The proposed method aims to provide fast and reliable road marking perception for high-definition mapping and intelligent transportation applications. Methods： LRMSNet adopts an encoder-decoder architecture and integrates convolutional neural networks, attention mechanisms, wavelet transform, and visual state space modeling into a unified framework. First, a lightweight bottleneck module is designed to capture multi-scale semantic information while reducing the number of model parameters. Second, an efficient visual state space pyramid is proposed to enhance global context modeling through multi-scale feature aggregation and linear-complexity computation. Third, a feature optimization module is introduced to refine skip connections between the encoder and decoder, thereby improving the preservation of spatial details and boundary information. These designs allow LRMSNet to simultaneously strengthen local detail representation and long-range dependency modeling under limited computational overhead. Results： Experiments are conducted on the public CamVid dataset and a self-collected Wuhan road marking dataset. LRMSNet achieves mean intersection over union scores of 68.89% and 63.11% on the two datasets, respectively. The model contains only 1.37 million parameters and requires 7.98 GFLOPs, demonstrating a lightweight computational profile. Compared with mainstream semantic segmentation models, LRMSNet substantially reduces the number of parameters while achieving superior segmentation performance, especially for fine-grained road markings. Conclusions： LRMSNet provides an efficient and accurate solution for road marking extraction from vehicle-borne images. By combining lightweight multi-scale representation, efficient global context modeling, and optimized feature fusion, the proposed network achieves a strong balance between segmentation accuracy and computational efficiency. This study offers practical technical support for rapid road marking extraction and has significant application potential in high-definition map generation, autonomous driving, and intelligent transportation systems.