To address the issues of low positioning accuracy and poor robustness in traditional visual simultaneous localization and mapping (SLAM) techniques within dynamic environments, an optimized ORB-SLAM3 algorithm incorporating YOLOv11n object detection was proposed. In the traditional system, an open neural network exchange (ONNX)-based YOLOv11n inference network was integrated to augment semantic information. Initial poses were generated using feature points from static regions, with map points then projected onto dynamic regions. A two-stage pose optimisation algorithm was integrated to filter static feature points and eliminate dynamic ones in dynamic regions, whereby pose estimation accuracy was improved and the number of high-quality feature points was increased. An additional thread was introduced beyond the original three, utilising pixels from keyframe regions to construct dense maps, providing rich environmental perception and understanding for subsequent human-computer interaction scenarios. Experimental results on the publicly available TUM dataset demonstrate that the proposed algorithm improves pose-estimation accuracy up to 98.3% relative to the baseline models. The proposed algorithm effectively mitigates the impact of dynamic objects on pose estimation while satisfying the requirements for dense map construction.