A multi-scale vehicle localization method was proposed to improve the localization accuracy of intelligent vehicles by constructing high definition maps from pavement landmarks. A high definition visual map was constructed based on pavement landmarks, where each landmark contains visual features, geometric structure information, and the positional coordinates in a reference coordinate system. Based on the constructed map, a coarse location of the vehicle was estimated through GPS matching, which was then improved by matching the visual features in the map to achieve landmark-level localization. Finally, the accurate and absolute position was achieved from the landmark geometry and its reference position, which thus could realize the multi-scale localization of the vehicle by high definition maps from pavement landmarks. A 3.4 km-long route of a university campus (including right-turn arrow, straight-way arrow, manhole, and so on) was taken as an example to construct a high definition visual map and then realize vehicle localization. Results showed that the mean and the max localization errors were 12.5 cm and 23.3 cm, respectively. The proposed mapping strategy based multi-scale localization method provides a new solution for intelligent vehicle localization.