A Data-Driven Estimation of Driving Style Using Deep Clustering

Accurately estimating driving style is crucial for designing personalized autonomous driving to enhance market acceptance. Focusing driving style estimation while driving, a novel model defined as deep clustering is proposed. Since the next generation simulation (NGSIM) dataset is complex and high-dimensional, a parameterized non-linear embedding from the original data space to a low-dimensional feature space by using deep neural networks (DNNs) is proposed to alleviate the "curse of dimensionality." We then propose a novel clustering layer to estimate the driving style of the encoded NGSIM data. Experimental results demonstrate that the NGSIM data divided into four groups shows better performance. Furthermore, compared with K-means, fuzzy C-means (FCM) and Gaussian mixture model (GMM), the proposed deep clustering model is capable of achieving superior performance in behavior analysis on public NGSIM dataset. Moreover, the deep clustering model has a stable performance on driving style estimation for different vehicle classes.