Unsupervised inference of driving behavior primitives considering feature correlations and temporal dynamics

As the smallest units of driving behavior, driving behavior primitive plays a crucial role in understanding driving dynamics. To better align the inferred driving behavior primitives with real-world driving scenarios, this paper proposes an unsupervised method consisting of two key stages: data segmentation and segments clustering. In data segmentation, a Hierarchical Bayesian Model-based Agglomerative Sequence Segmentation (H-BMASS) model is introduced to accurately capture the change points of variables across different dimensions, thereby avoiding under-segmentation issue that occurs when BMASS is applied to multivariate time series. In segments clustering, to address the limitation of LDA in capturing temporal dynamics in time series data, an Integrating Distribution- and Trend-based Latent Dirichlet Allocation (IDT-LDA) model is proposed. This model incorporates the change trend of data points to capture the temporal dynamics, determined by the derivatives and fluctuations. Compared to conventional BMASS, H-BMASS effectively identifies the change points of driving behavior, with clear structural differences between adjacent segments. Compared to the primitives obtained using Distribution-based LDA, the primitives clustered using IDT-LDA exhibit improved inter-class separability and enhanced interpretability. This unsupervised framework offers an effective method for inferring driving behavior primitives, which will benefit the development of AVs and ADASs.