Aiming at the problems of feature matching error and insufficiency of static features in dynamic environments, we proposed a dynamic point and line SLAM system based on point and line feature fusion and lightweight improved instance segmentation network. Three-dimensional line equations for line features were fitted by principal component analysis and spatial validation model are constructed to correct depth outliers. Secondly, the YOLOv8-seg backbone network is reconstructed by PP-LCNet and integrated a lightweight attention mechanism to improve the network inference speed while maintaining the detection accuracy. Thirdly, we integrated line segment midpoint and endpoint projection error, combined with the Gaussian residual model to construct a dynamic probability prediction mechanism for point and line features. Validated by the TUM RGB-D dataset, this system outperforms the benchmark algorithms such as ORB-SLAM3 in terms of absolute trajectory error and relative position error, in which the maximum reduction of ATE reaches 94.35%, which fully verifies the robustness of the proposed method and the enhancement of the positioning accuracy in the dynamic environment.