A method to improve the localization accuracy of intelligent vehicles was proposed by using map representation model from three-dimensional (3D) point cloud polarization. The point cloud polarization image was used as the node in this model, and the global position representation of the node could be realized through high-precision Global Positioning System (GPS) and Euler angle. The 2D and 3D features of the point cloud were then extracted from the polarization image to realize the multi-scale feature representation of the node, and the numerical description and virtual reconstruction of the road scene were realized through a series of polarization nodes. During the localization process, the 3D laser point cloud was in real-time acquired for polarization representation, and multi-scale feature matching was carried out with map nodes to realize the map localization of the intelligent vehicles. Specifically, map nodes were first filtered through GPS matching or topological localization based on the stability condition of the GPS signal of the intelligent vehicle, and a localization candidate set was obtained for the coarse localization. Then, the nearest map node was detected by matching the 2D point cloud features of the polarization image from the candidate set to complete the node localization. Finally, for the metric localization, the position of the intelligent vehicle was calculated by using the 3D point cloud feature matching results and the global position of the nearest map node. The experiment was carried out in two typical scenarios. The node localization accuracy rate was 98.7%, the average localization error was 21.4 cm, and the maximum localization error was 42.9 cm. Results show that the proposed algorithm has the advantages of high positioning accuracy, strong robustness, low cost, and simple calculation process.