The local driving behaviors varies between regions. Moreover, the statistical characteristics of surrounding vehicles' motions may significantly affect the performance of the policies of autonomous vehicles. It means autonomous vehicles (AV) driving policy may not work in a new area, which further limits the driving of AVs across regions. How to distinguish different scenarios from the perspective of affecting decision performance? This paper proposes the traffic scenario characteristics extractor and detector, using variational autoencoder (VAE). VAE is an unsupervised learning method, which can reconstruct the traffic data by the neural networks. This method uses vehicles state transition date as input and extract latent variables in two dimensions. In this way, the extracted hidden variables can represent the driving characteristics of the environment. Furthermore, the scenario characteristics detector relies on the similarity of the hidden variables, using KL-divergence. A policy can be used in the places has the similar characteristics. The method is tested by the NGSIM (Next Generation Simulation) and highD dataset. The VAE are trained One hundred thousand sets of data in ten minutes. The results indicate that this method can accurately distinguish between regions, and detect the similar traffic scenarios.

CICTP