<p>Mobile robots operating in dynamic environments often encounter challenges due to interference from moving objects, compromising the performance of their SLAM systems. To address this issue, we propose a depth camera based semantic-aware visual SLAM approach for mobile robots, aimed at dynamic scene tracking and reconstruction. By leveraging a lightweight Yolact++ network, we segment and detect objects in the scene, eliminate dynamic features through prior knowledge and epipolar constraints, and utilize high-confidence static features for robot’s pose estimation. Subsequently, the overall similarity between images is evaluated based on the similarity of the matched semantic object regions between images to select high-quality candidate keyframes, effectively enhancing the precision of loop closure detection. Furthermore, we employ Delta Generalized Labeled Multi-Bernoulli (Delta-GLMB) filtering to track the masks of prior dynamic objects provided by Yolact++. A global semantic octree map is then constructed, integrating both the semantic knowledge of the static scene and the tracking information of dynamic targets, which in turn realizes the tracking &amp; reconstruction of the dynamic scene. This allows mobile robots to mimic human behavior and cognitive abilities, not merely to construct a map of unknown environments more efficiently, but also to track the state of moving targets in the field-of-view(FoV) in real-time, enabling self-localization and navigation functions. Experimental results demonstrate the effectiveness of our approach in improving the robustness, accuracy, and adaptability of mobile robots in dynamic environments.</p>