Complementary Deep and Shallow Learning with Boosting for Public Transportation Safety

Shengda Luo, Alex Po Leung, Xingzhao Qiu, Jan Y. K. Chan, Haozhi HuangSensors 2020 · DOI

TL;DR

This paper introduces a novel boosting method that combines deep and shallow learning techniques to accurately predict unsafe driving behavior in public transportation using CAN bus data. By developing a new feature extraction method and evaluating on a real-world dataset, they achieve state-of-the-art performance, significantly improving road safety predictions.

What Problem Does This Paper Address?

Road traffic accidents are a major global concern, causing significant financial losses and human suffering. A substantial portion of these accidents are attributed to human error, particularly among public transportation drivers. While monitoring driver behavior is crucial for safety, existing methods for classifying safe vs. unsafe driving using vehicle sensor data (like CAN bus data) face challenges. These include high misclassification rates, a lack of informative features in raw CAN bus data, and the scarcity of publicly available, labeled datasets for this specific domain. This makes it difficult to build and evaluate accurate machine learning models for real-world road safety predictions.

Methods

The core of this paper lies in two novel contributions:

  • Feature Extraction: Recognizing that raw CAN bus data often lacks sufficient informative features, the authors propose a new method to extract richer time-series features. This involves calculating differences and gradients of existing features over specific time periods (e.g., two minutes) to capture dynamic aspects of driving behavior, such as acceleration, which are not directly recorded. This method is designed to be general and can be used with various machine learning classifiers.
  • Boosting Method: The authors introduce a heterogeneous boosting method that synergistically combines seven state-of-the-art machine learning algorithms: Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), Naive Bayes, Discriminant Analysis, Adaptive Boosting (AdaBoost), and Long Short-Term Memory (LSTM). This ensemble approach aims to leverage the strengths of each individual model to achieve superior classification performance compared to any single method. The boosting framework iteratively trains these "weak learners" and combines their predictions to form a strong final classifier.

Datasets

The study utilizes two datasets:

  • Warrigal Dataset: A publicly available dataset containing vehicle state information and vehicle-to-vehicle communications. For this study, a subset of data recorded on a specific day was used. Safety labels were derived by considering constant communication through wireless devices as potentially unsafe driving behavior.
  • TransMac Dataset: A newly collected, real-world dataset from TransMac, a public bus company in Macao. This dataset consists of CAN bus data from buses, with safety labels provided by domain experts. It contains 6451 records with 24 features, with 507 unsafe and 5944 safe cases. This dataset is particularly valuable as it addresses the lack of publicly available, labeled data for public transportation safety.

Key Insights and Ideas

  • Feature Engineering is Crucial: The paper strongly emphasizes that even with powerful machine learning models, the quality and richness of features are paramount. Their novel feature extraction method significantly boosts performance, highlighting the importance of domain-specific feature engineering for time-series data.
  • Ensemble Power: Combining diverse machine learning models (deep and shallow) through boosting proves more effective than relying on any single model. This suggests that different algorithms capture different aspects of the driving behavior data.
  • Real-World Data is Key: The creation and use of the TransMac dataset are critical. It validates the proposed methods in a practical, industry-relevant scenario, overcoming the limitations of synthetic or less specific datasets.
  • Time-Series Dynamics Matter: Capturing how features change over time (gradients, differences) is essential for understanding driving behavior, as static feature values alone are insufficient.

Key Results

  • The proposed boosting method, combined with the novel feature extraction technique, consistently outperformed seven other state-of-the-art machine learning methods on both the Warrigal and TransMac datasets.
  • On the TransMac dataset, the proposed method achieved a performance improvement of 5.9% and 5.5% over other methods when using 70% and 90% of the data for training, respectively, compared to using raw features.
  • The feature extraction method itself demonstrated a significant positive impact, improving the performance of all tested machine learning methods, including the proposed boosting approach.
  • The study found that the number of weak learners (parameter U) in the boosting method did not need to exceed 5 to achieve optimal performance, despite having seven potential base learners.

Limitations and Future Work

  • Computational Cost: The paper notes that boosting methods can be more computationally expensive during training compared to individual models, as indicated in Table 10.
  • Generalizability of Feature Extraction: While the feature extraction method is presented as general, further validation across different vehicle types and driving environments would be beneficial.
  • Real-time Implementation: The current study focuses on offline analysis. Future work could explore the real-time implementation of this system for immediate driver feedback and intervention.
  • Broader Safety Aspects: The study primarily focuses on driver behavior. Future research could integrate other factors contributing to road safety, such as environmental conditions or vehicle maintenance.

Summary updated

BibTeX
@article{luo2020complementary,
  author = {Shengda Luo and Alex Po Leung and Xingzhao Qiu and Jan Y. K. Chan and Haozhi Huang},
  title = {{Complementary Deep and Shallow Learning with Boosting for Public Transportation Safety}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20174671},
}

Loading related content...

Figure