Simultaneous localization and mapping (SLAM) is a core technology for the autonomous positioning and navigation of intelligent mobile robots. Dynamic SLAM has the problem of an insufficient number of features after removing dynamic point features, which leads to lower accuracy of pose estimation. Meanwhile, the time consumption of dynamic target detection based on deep learning is huge, which limits the practical application of dynamic SLAM algorithms in practice. To address the above problems, we proposed a dynamic point–line SLAM system (DPLS-SLAM) based on point–line feature fusion and lightweight improved YOLOv8seg instance segmentation network. The improved EDLines algorithm is used to extract high-quality line features, and a spatial consistency verification model is introduced, which uses RANSAC to fit 3D line equations and combines with Gaussian model to correct the line feature depth outliers, effectively solving the problem of the inconsistency between the depth recovery of line features and the scene texture. Secondly, a lightweight improved YOLOv8seg instance segmentation network is proposed to provide real-time semantic labels and object masks for dynamic object detection by reconfiguring the backbone network via EfficientNetV2 and adding a lightweight attention mechanism, which improves the inference speed with little change in accuracy. In addition, an adaptive weighted projection error model fusing line direction and distance information is constructed and combined with optical flow tracking to optimize the line feature matching efficiency. Experimental results on the TUM RGB-D dataset and real dynamic environments show that DPLS-SLAM outperforms existing dynamic SLAM solutions in both positioning accuracy and real-time performance. Compared with ORB-SLAM3, the absolute trajectory error is reduced by up to 94.35%, and the relative pose error is also reduced, verifying its effectiveness and robustness in dynamic environments.