HVAE-DC: A hierarchical variational autoencoder-based deep clustering model for multi-level driving behavior

Dongjian Song

Abstract: Unsupervised clustering of driving behavior serves as a critical foundation for driving data analysis, which is of great significance to enhancing autonomous vehicles' ability to comprehend complex environments and achieve high-performance driving. Existing methods often fail to simultaneously represent macro-level semantic categories (e.g., cruise, lane change) and micro-level operational intensity (e.g., emergency, normal), leading to oversimplified clustering results. To overcome this limitation, we propose a Hierarchical Variational Autoencoderbased Deep Clustering (HVAE-DC) model. The model constructs a hierarchical latent space with learnable probabilistic priors: macro-level priors capture semantic distributions of behaviors, while micro-level priors model continuous variations in operational intensity. A joint optimization strategy is designed to synchronously train hierarchical encoders via variational inference with KL divergence constraints, alongside dual supervision from reconstruction and clustering losses. This ensures macro-level variables focus on categorical features and micro-level variables encode intensity details, enabling multi-level behavior parsing from semantics to operational specifics. Experiments on the real vehicle-based Sag-DB dataset and the real-world AV2-DB dataset demonstrate that HVAE-DC outperforms comparative models in clustering performance and generalization ability, providing a solid foundation for autonomous driving data analysis.