Data-driven Semi-supervised Machine Learning for Detecting Abnormal Driving Behaviors using Surrogate Safety Measures

核心概念

This study develops a semi-supervised Hierarchical Extreme Learning Machines (HELM) model that leverages unlabeled data for self-supervised pre-training and partially labeled data for fine-tuning to accurately detect various abnormal driving behaviors, including sudden acceleration, rapid lane-changing, and close lane-changing. The study also introduces Surrogate Safety Measures (SSMs), specifically two-dimensional Time-to-Collision (2D-TTC), as important input features to enhance the detection performance.

要約

This study analyzes a large-scale real-world naturalistic driving dataset, the CitySim dataset, and identifies several types of abnormal driving behaviors, including:

Rapid acceleration and emergency brake behavior: Detected by analyzing extreme acceleration and deceleration points.
Rapid lane-changing behavior: Detected by analyzing outliers in lateral acceleration beyond a standard deviation threshold.
Close lane-changing behavior: Detected by analyzing the distance between vehicles during lane changes, with a distance less than 0.5 meters considered severe abnormal behavior.

The study then develops a semi-supervised Hierarchical Extreme Learning Machines (HELM) model that leverages unlabeled data for self-supervised pre-training and partially labeled data for fine-tuning to accurately detect these abnormal driving behaviors. Additionally, the study introduces Surrogate Safety Measures (SSMs), specifically two-dimensional Time-to-Collision (2D-TTC), as important input features to enhance the detection performance.
The results demonstrate that the proposed semi-supervised HELM model with 2D-TTC features outperforms other baseline models, achieving the best accuracy at 99.58% and the best F1-score at 0.9913. The ablation study further highlights the significance of incorporating SSMs, as the inclusion of 2D-TTC improves the detection performance by over 5% across all tested models.

統計

The study utilizes the CitySim dataset, which contains vehicle trajectory information extracted from videos captured by 12 drones, spanning six road geometry typologies, including freeway segments, signalized intersections, and stop-controlled junctions.
The dataset provides the following key features:

frameNum
carId
carCenterX (ft)
carCenterY (ft)
headX (ft)
headY (ft)
tailX (ft)
tailY (ft)
Speed (mph)
Heading (°)
laneId
From these raw features, the study derives additional features, including:

Longitudinal acceleration
Lateral acceleration
Inter-vehicle distance
Two-dimensional Time-to-Collision (2D-TTC)

引用

"Detecting abnormal driving behavior is critical for road traffic safety and the evaluation of drivers' behavior."
"Machine learning (ML)-based approaches have shown great promise in detecting abnormal driving behaviors. They can learn complex patterns, adapt to changing scenarios, handle large and diverse datasets, and detect unusual behaviors with optimized processes."
"Surrogate Safety Measures (SSMs) show promise as robust indicators and proxy measurements of sustainable road safety that can supplement or replace traditional historical crash analyses."

抽出されたキーインサイト

Data-driven Semi-supervised Machine Learning with Surrogate Safety Measures for Abnormal Driving Behavior Detection

by Lanxin Zhang... 場所 arxiv.org 04-11-2024

https://arxiv.org/pdf/2312.04610.pdf

深掘り質問

How can the proposed semi-supervised HELM model be extended to predict future abnormal driving behaviors, enabling proactive safety interventions?

In order to extend the proposed semi-supervised HELM model to predict future abnormal driving behaviors, a predictive modeling approach can be implemented. This would involve training the model on historical data to learn patterns and trends in abnormal driving behaviors. By incorporating time-series analysis techniques, the model can forecast potential abnormal driving behaviors based on the current driving conditions and past behavior patterns. This predictive capability would enable proactive safety interventions by alerting drivers, autonomous vehicles, or traffic management systems to potential risks before they occur. Additionally, real-time data streams can be integrated into the model to provide up-to-date information for more accurate predictions.

What counter-arguments could be made regarding the reliance on Surrogate Safety Measures (SSMs) for abnormal driving behavior detection, and how could the limitations of SSMs be addressed?

One counter-argument against the reliance on Surrogate Safety Measures (SSMs) for abnormal driving behavior detection is the potential for inaccuracies or biases in the data. SSMs are indirect indicators of safety and may not always capture the full complexity of driving behaviors. Additionally, SSMs may not be universally applicable across all driving scenarios or environments, leading to limitations in their effectiveness. To address these limitations, a multi-faceted approach can be adopted, combining SSMs with other data sources such as sensor data, video footage, and driver behavior analysis. By integrating multiple data streams and validation methods, the model can enhance the accuracy and reliability of abnormal driving behavior detection.

Given the success of the HELM model in this study, how could the framework be adapted to detect and classify different types of abnormal driving behaviors, such as distracted or impaired driving?

To adapt the HELM model to detect and classify different types of abnormal driving behaviors, such as distracted or impaired driving, additional features and data sources specific to these behaviors can be incorporated into the model. For distracted driving, features like eye movement tracking, smartphone usage patterns, and steering wheel movements can be included. Impaired driving detection can benefit from physiological data such as heart rate variability, blood alcohol levels, and erratic driving patterns. By expanding the feature set and training the model on diverse datasets that encompass various abnormal driving behaviors, the HELM framework can be tailored to classify and differentiate between different types of risky driving behaviors effectively.