Virtual navigation in 3D scenes requires both smooth and intelligent camera control, which is a very challenging task. In this paper, a real time navigation system that could automatically generate an effective and smooth camera flyby trajectory was presented. Firstly, offline geometric and semantic features of models were analyzed in the given 3D scene, including buildings, roads and other visible parts, to evaluate the importance of models. Secondly, the path from origin to destination was sampled adaptively, and visual images from different views on each sampled point were stored effectively. Finally, the best view on each sampled point was extracted and connected smoothly to get the camera flyby trajectory. The experiments done in four different 3D scenes clearly demonstrate the efficiency and intelligence of the proposed algorithm, and the user investigation also verifies its superiority over the existing methods.