[Paper Review] An Automated Machine Learning (AutoML) Method for Driving Distraction Detection Based on Lane-Keeping Performance

Chen Chai, Juanwu Lu+3 more|arXiv (Cornell University)|Mar 10, 2021

Human-Automation Interaction and Safety14 citations

TL;DR

This paper proposes AutoGBM, an automated machine learning (AutoML) framework that detects phone-related driving distractions using lane-keeping performance data. By integrating knowledge-based feature extraction, recursive feature elimination (RFE), and Bayesian optimization for hyperparameter tuning, AutoGBM achieves 80% accuracy at the group level and over 90% at the individual level, demonstrating high reliability and adaptability to driver-specific differences.

ABSTRACT

With the enrichment of smartphones, driving distractions caused by phone usages have become a threat to driving safety. A promising way to mitigate driving distractions is to detect them and give real-time safety warnings. However, existing detection algorithms face two major challenges, low user acceptance caused by in-vehicle camera sensors, and uncertain accuracy of pre-trained models due to drivers individual differences. Therefore, this study proposes a domain-specific automated machine learning (AutoML) to self-learn the optimal models to detect distraction based on lane-keeping performance data. The AutoML integrates the key modeling steps into an auto-optimizable pipeline, including knowledge-based feature extraction, feature selection by recursive feature elimination (RFE), algorithm selection, and hyperparameter auto-tuning by Bayesian optimization. An AutoML method based on XGBoost, termed AutoGBM, is built as the classifier for prediction and feature ranking. The model is tested based on driving simulator experiments of three driving distractions caused by phone usage: browsing short messages, browsing long messages, and answering a phone call. The proposed AutoGBM method is found to be reliable and promising to predict phone-related driving distractions, which achieves satisfactory results prediction, with a predictive power of 80\% on group level and 90\% on individual level accuracy. Moreover, the results also evoke the fact that each distraction types and drivers require different optimized hyperparameters values, which reconfirm the necessity of utilizing AutoML to detect driving distractions. The purposed AutoGBM not only produces better performance with fewer features; but also provides data-driven insights about system design.

Motivation & Objective

• Address low user acceptance of in-vehicle camera-based distraction detection by using non-invasive lane-keeping performance data.
• Overcome challenges in existing models related to individual driver differences and imbalanced datasets.
• Develop an automated, end-to-end pipeline that self-optimizes model selection, feature engineering, and hyperparameter tuning for distraction detection.
• Provide data-driven insights into driver monitoring system design by identifying key features and optimal model configurations per driver type.

Proposed method

• Employ knowledge-based feature extraction from lane-keeping performance data, including lateral deviation, steering angle, and their statistical and temporal features.
• Apply recursive feature elimination (RFE) to rank and select the most informative features, reducing dimensionality while preserving predictive power.
• Use Bayesian optimization to automatically tune hyperparameters of XGBoost, including learning rate, max depth, and L1/L2 regularization.
• Integrate algorithm selection and hyperparameter tuning into a single AutoML pipeline, enabling self-optimization of the entire modeling process.
• Leverage XGBoost as the base classifier due to its robustness, interpretability, and ability to handle noisy and missing data.
• Utilize weight-based and gain-based feature importance scores from decision trees to assess and rank feature relevance.

Experimental results

Research questions

• RQ1 Can an AutoML pipeline effectively detect driving distractions using only lane-keeping performance data, without relying on in-vehicle cameras?
• RQ2 How does the AutoML approach handle individual driver differences in distraction detection accuracy?
• RQ3 What are the optimal hyperparameter configurations for detecting different types of phone-related distractions (e.g., short message browsing, long message reading, call answering)?
• RQ4 To what extent does feature selection via RFE improve model performance and interpretability in distraction detection?
• RQ5 How does the AutoML system compare in performance across different driver profiles, such as young vs. older, risky vs. conservative drivers?

Key findings

• AutoGBM achieved 80% prediction accuracy at the group level and over 90% accuracy at the individual level across 28 drivers, demonstrating strong generalization and reliability.
• Different distraction types required distinct optimal hyperparameter settings, confirming the necessity of personalized model tuning via AutoML.
• Young and risky drivers were easier to detect for distraction than older, more conservative drivers, requiring fewer estimators and achieving higher accuracy.
• Detecting complex distractions like reading long messages required deeper models, more estimators, and stronger L1/L2 regularization to achieve high accuracy.
• The AutoML pipeline successfully reduced feature count while maintaining high performance, highlighting its efficiency and interpretability.
• The method revealed data-driven insights into driver monitoring system design, such as the importance of specific lane-keeping features and model configurations per driver type.