Decision-Oriented Driving Scenario Recognition Based on Unsupervised Learning

The statistical characteristics of surrounding vehicles' motions may significantly affect the performance of autonomous vehicles (AV). Therefore, AV driving policy may not work in a new area, which limits the driving of AVs across regions. This paper proposes a means of extracting and detecting traffic scenario characteristics, using a Variational Autoencoder (VAE). VAE is an unsupervised learning method, which can reconstruct the traffic data using neural networks. This method uses a vehicle's state transition date as input and extracts latent variables in two dimensions. The extracted hidden variables can represent the driving characteristics of the environment. The scenario characteristics detector relies on the similarity of the hidden variables, using KL-divergence. The method is tested by the NGSIM (Next Generation Simulation) and highD dataset. The VAE are trained on 100,000 sets of data in 10 minutes. The results indicate that this method can accurately distinguish between regions, and detect traffic scenarios.