Analysis of Road-User Interaction by Extraction of Driver Behavior Features Using Deep Learning
DOI: 10.1109/access.2020.2965940
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 analyzing complex interactions between road environments and driver behavior, a critical factor in road safety research. Traditional methods for extracting latent features from driving data, such as Principal Component Analysis (PCA) and Kernel PCA, are limited by their reliance on linear transformations or high computational costs, respectively. To overcome these limitations, the authors propose an improved unsupervised deep learning model using a Denoising Stacked Autoencoder (SDAE). The primary objective is to extract high-level driving behavior features from raw kinematic data and visualize them as a graphical representation, specifically an RGB-colored trajectory, to detect patterns of driving behavior, road complexity, and specific events. The experimental design involved collecting data from ten participants during real driving tests conducted in an industrial zone in Bologna, Italy. The tests utilized a Racelogic Video V-Box Pro device to record kinematic measures, including longitudinal and vertical speed, longitudinal and transversal acceleration, combined G-forces, and vehicle heading, at a 10Hz frequency. The driving circuit consisted of a 2500-meter route divided into a complex segment with high traffic and intersections and a simpler residential segment. Each participant completed two sessions, each comprising three laps: an adaptation lap, a baseline lap, and a test lap involving simulated pedestrian crossings to induce workload. The SDAE model processed these six kinematic inputs using a sliding window of 60 inputs, encoding them through hidden layers (40, 20, 10 neurons) into a three-dimensional output mapped to RGB colors. The model employed a hyperbolic tangent activation function, L2 regularization to prevent overfitting, and a sparse term to ensure feature clarity. The results demonstrated that the SDAE model effectively extracted latent features that outperformed PCA and Kernel PCA in terms of linear separability and consistency. The generated driving color maps successfully distinguished between different driving contexts and behaviors. Specifically, the complex road segments were consistently associated with red or yellow-green colors, while the simpler segments were associated with blue colors. Furthermore, the model validated its ability to reflect driver workload; subjective assessments using the NASA-TLX survey showed a significant decrease in mental workload between the first and second driving sessions. This reduction in cognitive load corresponded with observable changes in the color maps, confirming that the extracted features are predictive of the drivers' mental state. The visualization method allowed for the differentiation of basic driving behaviors, such as high-speed forward motion and acceleration changes, which previous methods failed to distinguish clearly. The significance of this work lies in its demonstration that unsupervised deep learning can effectively process high-dimensional, non-linear driving data to produce interpretable visualizations of driver-vehicle-environment interactions. By mapping latent features to an RGB color space, the method provides a novel tool for analyzing driving styles and detecting road complexity without requiring labeled data. The findings suggest that such models can support real-time safety advice and contribute to the development of autonomous driving systems by accurately predicting driver behavior patterns. The study confirms that SDAE is a robust alternative to traditional linear methods for extracting meaningful insights from naturalistic driving data.
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | Crossref | — | — | 1 | 2026-06-19 |
| archive | success | unpaywall | — | — | 2 | 2026-06-25 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-20 |
| chunk | success | chunk | — | — | 1 | 2026-06-20 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-20 |
| enrich | success | openalex | — | — | 1 | 2026-06-20 |
| promote | success | — | — | — | 1 | 2026-06-19 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-20 |
| verify | success | — | — | — | 1 | 2026-06-26 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
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