Extraction of driving behavior primitives considering driver expectation and vehicle dynamics.

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Transportation Engineering

Background:

• Fine-grained analysis of driving behavior is essential for developing advanced driver-assistance systems (ADAS) and autonomous vehicles (AVs).
• Existing methods often struggle to capture the hierarchical and nuanced nature of human driving patterns.

Purpose of the Study:

• To develop a novel framework for extracting hierarchical driving behavior primitives.
• To improve the accuracy and interpretability of driving behavior segmentation and clustering.

Main Methods:

• A two-stage framework combining Bayesian Model-based Agglomerative Sequence Segmentation (BMASS) for behavior segmentation and Variable Coupling-based Latent Dirichlet Allocation (VC-LDA) for primitive clustering.
• Construction of a multi-type feature space incorporating vehicle motion states and driver expectations.
• Development of a features-coupling-aware discretization process within VC-LDA to handle non-linear feature coupling and temporal asynchrony.

Main Results:

• The proposed BMASS model achieved precise behavior segmentation, with segment durations and positioning as key quality metrics.
• The VC-LDA method significantly outperformed GMM-LDA in clustering, demonstrating lower perplexity and higher intra-class compactness.
• The framework provides enhanced physical interpretability of driving states.

Conclusions:

• The novel framework offers an automated and efficient method for understanding and modeling driver behavior.
• This approach provides valuable insights for the development of safer and more sophisticated ADAS and AVs.
• The VC-LDA method represents a significant advancement in clustering driving states.