Traffic Accident Data Generation Based on Improved Generative Adversarial Networks

Zhijun Chen, Jingming Zhang, Yishi Zhang, Zihao HuangSensors 2021 · DOI

TL;DR

This paper tackles the critical issue of scarce traffic accident data by proposing an improved Generative Adversarial Network (GAN) model to generate synthetic accident data. Their novel GAN architecture effectively creates realistic traffic accident data, which, when used to augment existing datasets, significantly enhances the performance of traffic accident recognition models.

What Problem Does This Paper Address?

Traffic accidents pose a severe risk to human lives and property, making rapid recognition and warning crucial for mitigating their impact. However, a significant hurdle in developing robust traffic accident recognition systems is the scarcity and difficulty of collecting comprehensive accident data. Traditional machine learning models, especially deep learning approaches, often require large, diverse datasets to achieve high accuracy. The lack of sufficient accident data leads to models that are less reliable and prone to errors, particularly false alarms. This paper aims to bridge this data gap by generating synthetic traffic accident data that closely resembles real-world scenarios.

Methods

The core of this paper is the development of an Improved Generative Adversarial Network (GAN) specifically tailored for traffic accident data generation. Recognizing that standard GANs are often designed for image generation, the authors modified the generator network structure to be applicable to non-graphical data. Their proposed architecture features a multi-generator framework inspired by DenseNet, where multiple generators with multi-channel connections form the final generative network. This design aims to improve the generator's ability to capture complex data distributions.

The training process involves an adversarial game between the generator (G) and discriminator (D) networks. The generator learns to produce synthetic data that mimics the real data distribution, while the discriminator learns to distinguish between real and generated samples. To ensure stability and prevent gradient disappearance, the model incorporates batch normalization and uses the Adam optimizer. The authors also detail the specific network architectures for their generators (G1, G2, Gn) and discriminator, including activation functions (sigmoid and ReLU) and layer configurations.

Datasets

The study utilizes a dataset collected from urban beltways and highways in Wuhan, China, between June 4th and June 10th, 2018. This dataset comprises vehicle GPS data, including speed and acceleration information. After preprocessing and feature extraction, six key features were identified: average speed, variance of speed, standard deviation of speed, average acceleration, variance of acceleration, and standard deviation of acceleration. The dataset was divided into two categories: "Accident" and "Normal" traffic. For training, 800 samples were used, and 200 samples were reserved for testing. The data was normalized before being fed into the classifiers.

Key Insights and Ideas

  • GANs for Non-Graphical Data: The paper demonstrates the successful application of GANs for generating tabular, non-graphical data, specifically traffic accident features, which is a less explored area compared to image generation.
  • Multi-Generator Architecture: The novel multi-generator approach, inspired by DenseNet, appears to be a key innovation for improving the generator's capacity to learn and replicate complex data distributions.
  • Data Augmentation for Safety-Critical Tasks: The research highlights the significant impact of synthetic data generation on improving the reliability of safety-critical applications like traffic accident recognition, where data scarcity is a major bottleneck.
  • Statistical Validation is Crucial: The rigorous statistical testing (F-test and t-test) to confirm that generated data is not significantly different from original data provides strong evidence for the method's validity.

Key Results

The generated traffic accident data was found to be statistically similar to the original data, with p-values from F-test and t-test results generally exceeding 0.05, indicating no significant difference within a 95% confidence interval.

When the generated data was used to augment the original dataset for traffic accident recognition tasks, several classifiers showed improved performance:

  • Accuracy Increase: The maximum increase in accuracy for accident recognition was 3.05%.
  • False Positive Rate (FPR) Reduction: The maximum decrease in the false positive rate was 2.95%.

Specifically, the paper reports the following average recognition rates:

  • SVM: Original: 92.05%, Original + Generated: 92.7%
  • KNN: Original: 90.7%, Original + Generated: 91.15%
  • Ridge Regression: Original: 89.25%, Original + Generated: 90.5%
  • CNN: Original: 69.65%, Original + Generated: 72.7%

These results demonstrate that the augmented dataset significantly enhances the performance of these classifiers.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Data Scope: The collected data only includes vehicle-related accidents on urban beltways and highways. Accidents involving pedestrians and non-motor vehicles (like bikes or e-bikes) are not covered.
  • Model Optimization: There is still room for improvement in the generative and discriminative models. Further research could focus on optimizing the GAN architecture to enhance training efficiency and generative performance.

Future work suggested includes:

  • Expanding Application Scenarios: Extending the application to include non-motor vehicle or pedestrian-related traffic accidents.
  • Optimizing GAN Architecture: Further research into optimizing the generative adversarial neural network model for more efficient training and improved generative performance.

Summary updated

Abstract

For urban traffic, traffic accidents are the most direct and serious risk to people's lives, and rapid recognition and warning of traffic accidents is an important remedy to reduce their harmful effects. However, research scholars are often confronted with the problem of scarce and difficult-to-collect accident data resources for traffic accident scenarios. Therefore, in this paper, a traffic data generation model based on Generative Adversarial Networks (GAN) is developed. To make GAN applicable to non-graphical data, we improve the generator network structure of the model and used the generated model to resample the original data to obtain new traffic accident data. By constructing an adversarial neural network model, we generate a large number of data samples that are similar to the original traffic accident data. Results of the statistical test indicate that the generated samples are not significantly different from the original data. Furthermore, the experiments of traffic accident recognition with several representative classifiers demonstrate that the augmented data can effectively enhance the performance of accident recognition, with a maximum increase in accuracy of 3.05% and a maximum decrease in the false positive rate of 2.95%. Experimental results verify that the proposed method can provide reliable mass data support for the recognition of traffic accidents and road traffic safety.
BibTeX
@inproceedings{chen2021traffic,
  author = {Zhijun Chen and Jingming Zhang and Yishi Zhang and Zihao Huang},
  title = {{Traffic Accident Data Generation Based on Improved Generative Adversarial Networks}},
  booktitle = {Sensors},
  year = {2021},
  doi = {10.3390/s21175767},
}
Figure