STCNet: SVO-Guided Two-Stage Clustering Network for Driving Style Recognition

The classification of driving styles is of paramount importance in intelligent transportation systems (ITS) owing to its potential to improve driving safety, optimize driving comfort, and mitigate energy consumption. This paper introduces STCNet, an SVO-guided two-stage clustering network for driving style recognition, consisting of two main components: the representation module and the clustering module. Specifically, the representation module employs a long short-term memory (LSTM)-based autoencoder to derive a low-dimensional representation containing the driving behavior characteristics of both lateral and longitudinal driving data. The clustering module is divided into two stages: first, the fuzzy C-means clustering method is applied to estimate the membership of aggressive and conservative driving styles for initial rough clustering. Then, a social value orientation (SVO)-guided clustering is introduced to refine the clustering process and enhance the interpretability. Through a two-stage training process: pretraining the representation module and then jointly training these two modules, a high-quality and clustering-friendly representation is extracted, thereby improving clustering performance. The Next Generation Simulation (NGSIM) public dataset is leveraged to train and evaluate our approach, demonstrating its effectiveness in driving style recognition.