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

## TL;DR

This paper introduces a new framework to extract driving behavior primitives by combining driver expectations and vehicle dynamics, improving behavior analysis for autonomous systems.

## Contribution

A novel VC-LDA method is proposed for clustering driving behaviors with enhanced interpretability through feature coupling awareness.

## Key findings

- The VC-LDA model outperforms GMM-LDA with lower perplexity and better intra-class compactness.
- The framework improves behavior segmentation by integrating objective and subjective driving features.
- Segment durations and positioning metrics effectively assess segmentation quality.

## Abstract

Extracting meaningful driving behavior primitives is crucial for fine-grained behavior analysis. To develop primitives that more closely mirror the hierarchical nature of human driving, this paper introduces a novel framework and carries out on a 16 drivers’ dataset. The proposed method operates in two synergistic stages. First, for behavior segmentation, a multi-type feature space is constructed to capture both the objective vehicle motion states and the subjective driver expectations on vehicle performance. This space is then input into a Bayesian Model-based Agglomerative Sequence Segmentation (BMASS) model to achieve precise segmentation. Segment durations and the positioning of segmentation points serve as key metrics to assess the quality of behavior segmentation. Next, for primitive clustering, a Variable Coupling-based Latent Dirichlet Allocation (VC-LDA) method is proposed. The core innovation of VC-LDA lies in a features-coupling-aware discretization process. By considering the non-linear coupling and temporal asynchrony among features in the multi-type space, this process yields driving states with enhanced physical interpretability, providing a high-quality foundation for LDA clustering. Experimental results demonstrate that the VC-LDA model significantly outperforms the GMM-LDA, achieving a much lower perplexity and exhibiting higher intra-class compactness in topic-word distributions. This framework offers an automated and efficient pathway to understand and model driver behavioral patterns, providing valuable insights for the development of ADAS and AVs.