High-precision positioning, as a critical issue in intelligent vehicle research, holds significant practical implications for investigating low-cost, high-precision, and real-time positioning methods. This paper firstly utilizes the NetVLAD model fused with the convolutional neural network EfficientNet to extract global features from forward-looking road images, followed by dimensionality reduction using PCA. Secondly, the ORB algorithm is employed to extract local features from overhead road surfaces. Through manual measurement, common road signs on the overhead road surface are used to construct a three-dimensional world coordinate system, mapping the extracted 2D local feature points to 3D. Finally, combining different types of features to represent road node positions forms a high-precision visual node map. In the positioning stage, a multi-scale positioning strategy is adopted, proceeding from coarse to fine. During the initial positioning stage, effective map nodes are rapidly selected through GPS location information matching. In the node-level positioning stage, a Bayesian model is used to fuse global features and vehicle motion information to determine the nearest map positioning node. In the metric-level positioning stage, spatial position relationships between intelligent vehicles and the nearest map nodes are computed using the LM-optimized PNP algorithm based on matched feature points, completing the entire positioning process. Experimental results demonstrate that the proposed method achieves an average positioning accuracy of 95.2%, with an average error of approximately 28.84 cm, and an overall positioning time of around 315 ms.