With the rapid advancement of the low-altitude economy, the demand for high-precision positioning in urban environments has become increasingly pronounced. Traditional satellite navigation and wireless signal-based positioning methods are prone to issues such as occlusion and electromagnetic interference in complex urban scenarios. Meanwhile, visual positioning approaches relying solely on single geometric features (points, lines, surfaces) suffer from insufficient robustness in dynamic occlusion, lighting variations, and low-texture regions. We propose a visual positioning algorithm based on multi-level features, aiming to enhance positioning accuracy and stability in complex urban environments by integrating geometric and semantic features. Our algorithm commences with image preprocessing to extract point and line geometric features, then employs the YOLO v5 model for semantic segmentation to identify dynamic objects (e.g., pedestrians, vehicles) and eliminate their associated geometric features. By combining semantic prior knowledge with geometric characteristics, the algorithm effectively mitigates dynamic interference and strengthens scene comprehension. Experimental results demonstrate that, compared to algorithms such as ORB-SLAM2 and Manhattan-SLAM, our method achieves significantly higher positioning accuracy in dynamic scenarios of the TUM dataset, validating the efficacy of the multi-level feature fusion strategy. Future research will concentrate on enhancing semantic understanding and refining dynamic feature filtering in high-dynamic environments to further improve the algorithm’s robustness.