<p>Simultaneous localization and mapping (SLAM) is a core technology enabling intelligent robots to autonomously navigate unknown environments. However, dynamic objects in real-world environments introduce many unstable dynamic features into the SLAM process, making it difficult for traditional visual SLAM methods based on static scene assumptions to ensure accurate camera pose estimation and consistent map construction in dynamic scenarios. To overcome this challenge, we propose SGL-SLAM, a novel RGB-D visual SLAM system designed for dynamic environments, built upon the ORB-SLAM2 framework. The proposed system uses a combination of object detection and segmentation models to identify and remove typical dynamic targets (e.g., humans), while also applying geometric constraints to filter out unstable feature points caused by non-prior dynamic objects, significantly enhancing tracking accuracy and map stability. When dynamic objects occupy a large proportion of the scene, the number and quality of successfully matched point features significantly decrease after dynamic filtering, making the system prone to accuracy degradation and even tracking failure. To address this issue, the system incorporates line features as complementary geometric cues to assist point-based tracking and mapping, and employs an adaptive joint optimization strategy to improve robustness in dynamic environments. Experimental results demonstrate that SGL-SLAM achieves localization accuracy similar to or slightly superior than state-of-the-art dynamic SLAM methods on the TUM and Bonn datasets, confirming its effectiveness and practicality in dynamic environments.</p>