Driving Style Recognition under Connected Circumstance Using a Supervised Hierarchical Bayesian Model

Chen, Depeng; Chen, Zhijun; Zhang, Yishi; Qu, Xu; Zhang, Mingyang; Wu, Chaozhong · 2021 · Crossref

DOI: 10.1155/2021/6687378

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of recognizing individual driving styles to improve passenger satisfaction in automated vehicles. While automated driving systems are advancing, mismatched driving styles between drivers and passengers can lead to discomfort. Existing methods often rely solely on vehicle motion data or unsupervised learning, neglecting the significant influence of traffic context and prior expert knowledge. To overcome these limitations, the authors propose a supervised hierarchical Bayesian model that integrates driving context with motion data to uncover latent driving styles. The methodology employs a two-step process: data encoding and model training. First, the authors developed a Motion and Context Aggregation Model (MCAM) to encode raw driving data into discrete behavioral "words." This encoding transforms continuous variables—such as speed, acceleration, turn signals, lane offset, and radar-detected obstacles—into categorical features. For instance, acceleration is divided into five intensity levels, and obstacles are categorized by position and distance. Each driving trip is treated as a "document" composed of these behavioral words. Second, the authors applied a Labeled Latent Dirichlet Allocation (LLDA) model, a supervised topic modeling technique. Unlike traditional unsupervised LDA, LLDA incorporates prior knowledge by using expert-labeled driving styles (aggressive, moderate, and careful) as constraints. This allows the model to learn the distribution of latent driving styles while respecting known labels, thereby enhancing classification accuracy. The experimental validation utilized the Safety Pilot Model Deployment (SPMD) dataset, collected by the University of Michigan Transportation Research Institute. This dataset includes high-resolution vehicle trajectories, basic safety messages, and radar data from onboard and roadside devices. The authors preprocessed the data by removing invalid records and normalizing continuous features, such as capping speed at 30 meters per second. The encoded behavioral sequences were then input into the LLDA model, which used Gibbs sampling to estimate the proportion of each driving style for individual drivers. The results demonstrate that the proposed MCAM + LLDA framework effectively uncovers latent driving styles and shows strong agreement with real-world driving situations. By incorporating traffic context and supervised labels, the model provides a more nuanced understanding of driving behavior than methods relying solely on motion data. The study concludes that this approach offers theoretical guidance for recognizing driving behaviors, which can facilitate better individualized experiences in automated driving systems by reducing mismatches between driver styles and passenger preferences.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software, dataset resource
• Theoretical Contribution: computational model