Hengbo Ma, Franklin Zhao, Jessica Leu, and Yujun Zou
May 2018
Prediction and recognition in complex situation have significant influence on the overall performance of autonomous driving systems. Many works focusing on single driver’s behavior have been done. However, modeling multi-driver interaction, which is a more general case, is harder. In our project, we first divide human driving behavior into a hierarchical model which contains decision-making phase and maneuver phase. Next, we use classifiers to find the drivers’ high-level intention, i.e., decision making, and then, we use Gaussian Mixture Models to capture different human driving behavior given their high level decisions. Last, base on the assumption that decisions in training data are unknown to us, we use a variational autoencoder to learn the representation of different driver behavior models in latent space and make prediction accordingly. A simulation data set from ramp-merging scenario is used to verify each models.