Shareable Driving Style Learning and Analysis With a Hierarchical Latent Model

Lijun Sun (孙立君), Chaopeng Zhang, Wenshuo Wang, Zhaokun Chen, Jian Zhang, Lijun Sun, Junqiang Xi

TL;DR

This paper introduces a novel hierarchical latent model to understand and analyze driving styles, moving beyond fixed individual labels to a probabilistic framework where driving behavior is a mixture of shareable latent styles. This approach allows for capturing both within- and between-driver style similarities, offering a more nuanced understanding of human driving.

What Problem Does This Paper Address?

Current approaches to understanding driving styles often treat them as fixed, individual attributes. However, real-world driving behavior is dynamic and influenced by various factors like personality, environment, and immediate goals. This leads to inconsistencies within a single driver's behavior and raises questions about how driving styles are shared across different individuals. The paper aims to bridge this gap by developing a model that can learn and analyze these shareable driving styles, recognizing that a driver's behavior is a probabilistic combination of underlying, common driving styles. This is crucial for developing more adaptive and personalized intelligent vehicle systems.

Methods

The core of the paper is a hierarchical latent model designed to capture the complex relationship between driving behavior and shareable driving styles. The methodology involves several key steps:

  1. Behavior Fragmentation and Dimension Reduction: Raw, high-dimensional sensory data (speed, acceleration, yaw rate) is segmented into fixed-length fragments. For each fragment, statistical features (mean, max, min) are extracted. These statistics are then subjected to factor analysis to reduce their dimensionality into a smaller set of "common factors" (specifically, three factors: lateral, acceleration, and speed). This low-dimensional representation is termed "driving words."
  2. Hierarchical Latent Model:
    • Driving Style: Driving styles are treated as unobservable, latent variables. They are modeled as discrete categories (e.g., calm, normal, aggressive) using a categorical distribution. Crucially, these styles are assumed to be temporally self-shareable within a driver and shareable between drivers. The latter is modeled using a Dirichlet distribution for the parameters of the categorical distribution, allowing for shared information about driving style proportions across individuals.
  3. Driving Behavior: Driving behavior fragments are generated based on the learned common factors ("driving words"). This generation process is modeled using a categorical distribution parameterized by the latent driving style.
  4. Hierarchical Structure: The model combines these components hierarchically. The Dirichlet-Category and Dirichlet-Multinomial conjugations are used to create a computationally and algebraically convergent model that links driving behavior fragments to latent driving styles and then to individual drivers and their style mixtures.
  5. Model Learning: A collapsed Gibbs sampling approach is employed to infer the posterior distribution of the latent driving styles.
  6. Evaluation: The model's performance is evaluated using perplexity and Shannon entropy, metrics that quantify the uncertainty and semantic clarity of the learned driving styles, respectively. Subjective-objective consistency verification is also performed by comparing the model's output with expert and participant ratings.

Datasets

The study utilizes naturalistic driving data collected from 100 drivers in Changchun, China. The data was acquired using a Hongqi E-HS9 testing vehicle equipped with a CAN bus and a front-view camera, capturing longitudinal velocity, acceleration, and yaw rate at 100 Hz. The driving routes included both urban roads (14.4 km) and highways (7.6 km), covering various driving scenarios. Participants also completed a driving style questionnaire.

Key Insights and Ideas

  • Driving Behavior as a Mixture of Styles: The fundamental insight is that an individual's driving behavior isn't a single, fixed style but rather a probabilistic mixture of underlying, shareable latent styles. This moves beyond simplistic classifications.
  • Shareability is Key: The model explicitly captures the idea that driving styles are not isolated to individuals but are shared across drivers and can manifest differently over time and space for the same driver.
  • Hierarchical Structure for Nuance: The hierarchical latent model effectively disentangles the complex relationships between raw data, driving words, latent styles, and individual drivers, providing a structured way to analyze these interactions.
  • Personality Influences Style Proportions: The analysis reveals statistically significant correlations between personality traits (gender, age, driving experience) and the proportions of different driving styles, offering insights into why certain styles are more prevalent in specific demographic groups.
  • Environmental Adaptability: Driving styles, while habit-driven, can adapt to environmental cues like speed limits and road signs, demonstrating a dynamic interplay between ingrained habits and situational awareness.

Key Results

  • Subjective-Objective Consistency: The model achieved high accuracy (93% for urban, 94% for highway) in classifying driving styles against subjective ratings, demonstrating its effectiveness in capturing real-world driving behavior.
  • Aggressive Drivers are Distinctive: The aggressive driving style was found to be more easily identifiable, with high precision and recall, likely due to its distinct behavioral patterns (e.g., high speed, large acceleration).
  • Females Drive Calmer: The study found that female drivers exhibit a significantly lower proportion of aggressive styles compared to male drivers.
  • Young Drivers are More Aggressive: Drivers aged 21-30 showed a higher percentage of aggressive styles than older age groups.
  • Experience Matters: Drivers with less than five years of experience tended to be more conservative (calm style), with this proportion decreasing as experience increased.
  • Normal Style is Common: For drivers labeled as "normal," the "normal" driving topic was the most prevalent, indicating a self-consistent labeling.

Limitations and Future Work

  • Fixed Sampling Step: The adoption of a fixed sampling step for behavior fragmentation was for computational efficiency but might overlook the dynamic structure of driving behavior over time.
  • Tedious Testing for Driving Words: Determining the discrete "driving word" set requires extensive testing.
  • Future Work: The authors plan to incorporate factor analysis and temporal influences directly into the hierarchical latent model in a continuous space, moving away from discrete "driving words." They also aim to explore more sophisticated methods for handling the dynamic nature of driving behavior.

Summary updated

Abstract

Driving style is usually used to characterize driving behavior for a driver or a group of drivers. However, it remains unclear how one individual’s driving style shares certain common grounds with other drivers. Our insight is that driving behavior is a sequence of responses to the weighted mixture of latent driving styles that are shareable within and between individuals. To this end, this paper develops a hierarchical latent model to learn the relationship between driving behavior and driving styles. We first propose a fragment-based approach to represent complex sequential driving behavior in a low-dimension feature space. Then, we provide an analytical formulation for the interaction of driving behavior and shareable driving styles through a hierarchical latent model. This model successfully extracts latent driving styles from extensive driving behavior data without the need for manual labeling, offering an interpretable statistical structure. Through real-world testing involving 100 drivers, our developed model is validated, demonstrating a subjective-objective consistency exceeding 90%, outperforming the benchmark method. Experimental results reveal that individuals share driving styles within and between them. We also found that individuals inclined towards aggressiveness only exhibit a higher proportion of such behavior rather than persisting consistently to be aggressive.
BibTeX
@article{2024shareable,
  author = {Lijun Sun (孙立君) and Chaopeng Zhang and Wenshuo Wang and Zhaokun Chen and Jian Zhang and Lijun Sun and Junqiang Xi},
  title = {{Shareable Driving Style Learning and Analysis With a Hierarchical Latent Model}},
  journal = {IEEE Trans. ITS},
  year = {2024},
  doi = {10.1109/tits.2024.3374771},
}
Figure