EM-LSD: A lightweight and efficient model for multi-scale line segment detection

To address the challenges of detecting line segments in dynamic and geometrically complex environments, EM-LSD, a lightweight and efficient line segment detection model, is introduced. Accurate and efficient detection of line segments is critical for tasks such as environmental modeling and localization in SLAM, where the failure to extract robust line features can result in unreliable mapping and trajectory estimation. The design of EM-LSD is guided by the limitations of existing methods: traditional approaches often fail to capture multi-scale and global features in noisy scenes, while deep learning models with multi-stage architectures impose high computational costs, making them unsuitable for real-time applications. Inspired by the observation that multi-scale feature extraction is essential for handling diverse geometric structures, EM-LSD incorporates a Dense Atrous Convolution (DAC) module to effectively capture multi-scale information with minimal computational overhead. Additionally, the need for robustness against structural complexities and noise led to the integration of dual decoders with a Channel-Spatial Multi-scale Attention (CSMA) module and a Multi-scale Atrous Deformable Block (MADB), enabling adaptive feature representation. Experimental results on the Wireframe and YorkUrban datasets validate EM-LSD’s superior accuracy, robustness, and real-time performance, emphasizing its capability to support resource-constrained SLAM applications. This model not only addresses the limitations of existing methods but also enhances the reliability of environment modeling and localization, offering inspiration for the development of lightweight and efficient detection frameworks.