Machine Learning-based Fusion Convolutional Neural Network Approaches for Driver Injury Severity Prediction Using Highway Single-Vehicle Crash Data in RITI Communities

It was reported that more than 36,000 people lost their lives, 4.5 million people were injured, and 24 million vehicles were damaged in motor vehicle crashes in the United States in 2018. The economic costs of these crashes totaled $242 billion accounting for 1.6% of the U.S. gross domestic product. Frequent crashes with severe injuries and fatalities become more problematic in Rural, Isolated, Tribal, or Indigenous (RITI) communities. Significant research efforts have been devoted to developing better approaches to formulate driver injury severities and their impact factors in the past decades. For example, binary discrete models, such as binary logit and probit models, have been applied in many early studies. Many variations were then proposed to overcome the limitations of traditional binary logit and probit models, such as single injury outcome and unobserved effects of impact factors. Besides these statistical techniques, some researchers also applied machine learning approaches in traffic crash analyses. Abdelwahab et al. have applied multi-layer neural network models for vehicle injury severity classification. Li et al. predicted motor vehicle crashes using support vector machine models. Recent advances in artificial intelligence provided an opportunity to formulate multi-hidden-layer learning structures, i.e., deep neural network (DNN), which is capable to learn effective representations of data within unstructured data [8]. DNN approaches have been extensively studied in transportation research, such as short-term traffic flow prediction, traffic demand estimation, and traffic crash forecast. In this study, the research team proposes a fusion convolutional neural network model with random term (FCNN-R) for driver injury severity analyses. More specifically, the team plans to deploy fusion convolutional neural networks to investigate the relationships between impact factors and driver injury severities in RITI communities. The unobserved heterogeneity across different crash records is illustrated using a random error term with zero mean. A modified pseudo elasticity analysis is applied to uncover the essentiality of each variables for driver injury severities. The research findings are helpful for transportation agencies to develop cost-effective countermeasures to mitigate rural crash severities and minimize the rural crash risks and severities in the States of Alaska, Washington, Idaho, and Hawaii.

Language

• English

Project

• Status: Completed
• Sponsor Organizations:

  Office of the Assistant Secretary for Research and Technology

  University Transportation Centers Program
  Department of Transportation
  Washington, DC  United States  20590
• Managing Organizations:

  Center for Safety Equity in Transportation

  University of Alaska Fairbanks
  Fairbanks, AK  United States  99775
• Principal Investigators:

  Zhang, Guohui

  Prevedouros, Panos

  Ma, David

• Start Date: 20200801
• Expected Completion Date: 0
• Actual Completion Date: 0

Subject/Index Terms

• TRT Terms: Communities; Crash data; Crash severity; Drivers; Injury severity; Machine learning; Mathematical prediction; Neural networks; Rural areas; Single vehicle crashes; Traffic safety
• Geographic Terms: Alaska; Hawaii; Idaho; Washington (State)
• Subject Areas: Data and Information Technology; Highways; Safety and Human Factors; Society;

Filing Info

• Accession Number: 01748591
• Record Type: Research project
• Source Agency: Center for Safety Equity in Transportation
• Files: UTC, RIP
• Created Date: Aug 25 2020 5:29PM