A long short-term memory-based framework for crash detection on freeways with traffic data of different temporal...

Area of Science:

• Intelligent Transportation Systems
• Deep Learning for Traffic Safety
• Machine Learning Applications in Transportation

Background:

• Current traffic crash detection methods using machine learning struggle to simulate evolving traffic conditions and lack multi-resolution temporal data analysis.
• Existing models often use single temporal resolutions, failing to capture comprehensive traffic trends crucial for accurate crash risk assessment.
• Limitations in simulating dynamic traffic transitions and incorporating diverse temporal data hinder the effectiveness of current intelligent transportation systems.

Purpose of the Study:

• To propose a novel Long short-term memory (LSTM) based framework, LSTMDTR, that integrates traffic data from multiple temporal resolutions for enhanced crash detection.
• To address the limitations of existing models in simulating pre-crash traffic dynamics and utilizing varied temporal traffic data.
• To improve the accuracy and transferability of traffic crash detection systems within intelligent transportation frameworks.

Main Methods:

• Developed a Long short-term memory (LSTM) based framework (LSTMDTR) utilizing three LSTM networks, each processing traffic data at different temporal resolutions.
• Integrated outputs from the three LSTM networks using a fully-connected layer and employed a dropout layer to mitigate overfitting and enhance prediction performance.
• Implemented and validated the LSTMDTR model on real-world traffic datasets from California freeways (I880-N and I805-N).

Main Results:

• The LSTMDTR model achieved a highest crash accuracy of 70.43% on crash detection tasks.
• Demonstrated significant transferability, achieving 65.12% crash accuracy when models trained on one freeway were applied to similar freeways.
• Outperformed traditional machine learning methods and single/dual-resolution LSTM models in both crash detection accuracy and transferability.

Conclusions:

• The LSTMDTR framework effectively captures long-term dependencies and dynamic transitions in pre-crash traffic conditions using multi-resolution data.
• The model shows promising performance and transferability, offering a more robust solution for intelligent transportation systems aiming to improve road safety.
• Further research should focus on optimizing the number of neurons for balancing performance and computation time, and the dropout technique enhances generalization.