Three-dimensional (3D) city visualization is fundamental for information display in smarter cities. Existing 3D city visualization systems are mainly limited in two aspects. First

the data model of a 3D city is unsuitable for handling massive building data and does not support real-time network transmission. Second

given that the whole city is rendered in a homogeneous graphic style and that most buildings are highly similar in texture

structure

and height

the rendering results usually exhibit high visual complexity. In this study

we propose a technique for online 3D city visualization on the basis of human visual perception and spatial cognition theory. During preprocessing

our system creates a multiple representation of buildings on the basis of the building generalization technique. Then

at runtime

the system can dynamically choose an appropriate level of buildings on the basis of human interaction. The algorithm is tested on several large 3D city models. Experiment results confirm the applicability of our algorithm to interactive visualization and information intensity reduction of a 3D city. For the Leverkusen city model with 5 530 buidings

we have improred the frame rate to 19.4 frame/s. The proposed method can improve the efficiency of 3D city rendering and thus enable users to obtain relevant information through efficient interaction.