Driving Style Analysis Using Primitive Driving Patterns With Bayesian Nonparametric Approaches

Wenshuo Wang, Junqiang Xi, Ding ZhaoIEEE Trans. ITS 2019 · DOI

TL;DR

This paper introduces a novel framework for analyzing driving styles by learning "primitive driving patterns" directly from naturalistic driving data using Bayesian nonparametric methods. This approach overcomes the need for pre-defined pattern counts and allows for a more semantically rich understanding of individual driving behaviors and their variations.

What Problem Does This Paper Address?

Understanding a driver's "style" is crucial for various applications, including intelligent vehicle control, driver assistance systems, and even vehicle calibration. Previous methods often relied on statistical metrics (like mean and standard deviation of speed or acceleration) or manually defined patterns. While these approaches can capture static habits, they struggle to represent the dynamic, sequential nature of driving decisions and behaviors. Furthermore, they often require prior knowledge of the number of distinct driving patterns, which is not always feasible or accurate. This paper aims to address these limitations by developing a data-driven method that can automatically discover and semantically interpret primitive driving patterns without pre-specifying their quantity.

Methods

The core of the proposed framework is a Bayesian nonparametric approach that leverages a Hierarchical Dirichlet Process (HDP) combined with a Hidden Semi-Markov Model (HSMM).

  • HDP-HSMM: This model is designed to learn an unknown number of latent states (primitive driving patterns) from time-series data. The HDP allows the model to infer the optimal number of patterns, while the HSMM accounts for the variable duration of these patterns. This is a key departure from traditional Hidden Markov Models (HMMs) which assume fixed state durations and a known number of states.
  • Primitive Driving Patterns: These are defined as the fundamental building blocks of driver behavior. The paper extracts these patterns from three key variables: relative range (distance to the lead vehicle), relative range rate (speed difference), and subject vehicle acceleration.
  • Semantic Labeling: After learning the primitive patterns, they are semantically labeled. This involves:
  • Variable Segmentation: Each of the three feature variables is divided into discrete levels based on human perception thresholds (e.g., "close distance," "normal distance," "long distance" for range).
  • K-Means Clustering: The learned primitive patterns are then clustered using K-means. This helps to group similar patterns and assign semantic labels (e.g., "keeping distance by aggressive deceleration in a long distance").
  • Driving Style Analysis: Individual driving styles are characterized by the normalized frequency distribution of these semantically labeled primitive patterns. Differences between drivers are quantified using the Kullback-Leibler (KL) divergence between their pattern distributions.

The paper also compares the HDP-HSMM to other Bayesian nonparametric models like HDP-HMM and sticky HDP-HMM, demonstrating the superiority of HDP-HSMM in capturing realistic driving dynamics.

Datasets

The study utilizes the Safety Pilot Model Deployment (SPDM) database, which contains naturalistic driving data from 2,842 vehicles in Ann Arbor, Michigan. For this research, data from 18 equipped vehicles (representing 18 drivers) were used to train and evaluate the models. The data was collected at 10 Hz and included variables such as speed, steering angle, and acceleration/brake pedal position.

Key Insights and Ideas

  • Unsupervised Discovery of Driving Patterns: The most significant insight is the ability of the HDP-HSMM to automatically discover a meaningful number of primitive driving patterns without prior human specification. This is a major step towards truly data-driven driving behavior analysis.
  • Semantic Interpretation of Driving: By combining learned patterns with human perception thresholds, the framework provides a semantically interpretable understanding of driving behaviors. This moves beyond raw statistical data to describe what a driver is doing and why (in terms of their immediate driving context and actions).
  • HSMM for Realistic Durations: The use of HSMM is crucial for capturing the fact that drivers don't instantaneously switch between behaviors. Primitive patterns have varying durations, and HSMM models this more accurately than traditional HMMs.
  • KL Divergence for Inter-Driver Comparison: Using KL divergence to compare the distribution of primitive driving patterns offers a robust way to quantify similarities and differences in driving styles, providing a more nuanced view than simple statistical comparisons.

Key Results

  • Model Performance: The HDP-HSMM model demonstrated superior performance in terms of log-likelihood for both training and predictive log-likelihood on test data compared to HDP-HMM and sticky HDP-HMM. This indicates its better ability to segment and model car-following sequences.
  • Pattern Segmentation: The HDP-HSMM was shown to produce more realistic segmentations, avoiding the overly frequent and short-lived patterns observed with the sticky HDP-HMM, and capturing more nuanced behaviors than the HDP-HMM.
  • Semantic Labeling Success: The framework successfully segmented and labeled primitive driving patterns, allowing for the description of behaviors like "rapidly falling behind the lead vehicle by gentle accelerating in a normal distance."
  • Driving Style Characterization: Analysis of normalized frequency distributions revealed distinct driving preferences among drivers, particularly in how they manage distance, acceleration, and deceleration under different driving conditions (close, normal, long distance). For instance, some drivers consistently prefer to fall behind aggressively, while others prioritize maintaining a constant headway.
  • Inter-Driver Differences: KL divergence analysis highlighted significant differences in driving styles among drivers, with certain drivers showing unique preferences that set them apart from the group.

Limitations and Future Work

  • Data Limitations: The study used data from 18 drivers. While this is a reasonable sample size for demonstrating the method, a larger and more diverse dataset would enhance the generalizability of the findings.
  • Generalizability of Primitive Patterns: The paper focuses on car-following scenarios. Extending the framework to other driving maneuvers (e.g., lane changing, merging) would be a valuable future direction.
  • Subjectivity in Thresholds: While thresholds were based on human perception, there's an inherent subjectivity. Future work could explore more adaptive or data-driven methods for setting these segmentation boundaries.
  • Real-time Application: The current framework is primarily for offline analysis. Adapting it for real-time driving style recognition would be a significant practical advancement.

Summary updated

BibTeX
@article{wang2019driving,
  author = {Wenshuo Wang and Junqiang Xi and Ding Zhao},
  title = {{Driving Style Analysis Using Primitive Driving Patterns With Bayesian Nonparametric Approaches}},
  journal = {IEEE Trans. ITS},
  year = {2019},
  doi = {10.1109/tits.2018.2870525},
}
Figure