Safety Assessment for Vulnerable Road Users Using Automated Data Extraction with Machine-Learning Techniques

Date of Award

8-2024

Degree Name

Doctor of Philosophy

Department

Civil and Construction Engineering

First Advisor

Jun-Seok Oh, Ph.D.

Second Advisor

Valerian Kwigizile, Ph.D.

Third Advisor

Ron Van Houten, Ph.D.

Fourth Advisor

Osama Abudayyeh, Ph.D.

Keywords

Crash diagrams, deep learning, image processing, machine learning, safety, vru

Abstract

Road users at high risk of crash involvement are classified as Vulnerable Road Users (VRUs) due to their limited ability to respond in critical situations and lack of adequate protection. This has prompted transportation researchers and urban planners to prioritize their safety due to the increased incidence of injuries and fatalities. Enhancing our understanding of crashes involving VRUs is essential for identifying critical factors and devising effective countermeasures. This endeavor requires access to detailed, comprehensive, and impartial data. Crash report databases serve as a fundamental data source for analyzing such incidents. Detailed data on crashes involving VRUs are crucial for safety studies, as comprehensive descriptions of crash scenes and user behavior are predominantly found within crash narratives and diagrams. However, extracting and applying this information from traffic crash reports poses significant challenges. Therefore, an automated approach is needed to assist researchers in extracting crash patterns involving VRUs from traffic crash narratives and diagrams, facilitating safety analysis and the development of countermeasures.

This research has developed an image-processing framework that integrates machine-learning techniques to assess and enhance VRUs' safety. The practical application of this approach is in identifying and extracting unstructured hidden features from crash diagrams involving VRUs. Deep Convolutional Neural Networks (CNNs) are recognized for acquiring significantly more resilient and nuanced features than manually engineered ones. These features have demonstrated effectiveness across various computer vision applications, including but not limited to object detection, pattern recognition, and image classification. Different features are extracted using multiple CNN architectures trained on a diverse representation of crash diagram databases involving VRUs.

In the first part of the study, a framework is developed to identify and extract latent features from pedestrian crash data in Michigan. This part evaluates the performance and effectiveness of three deep CNN architectures—VGG-19, AlexNet, and ResNet-50—in classifying multiple features in pedestrian crash diagrams. These features include intersection type (3-leg or 4-leg), road type (divided or undivided), the presence of marked crosswalk, intersection angle (skewed or unskewed), the presence of Michigan left turn, and the presence of nearby residential area. The findings underscore the importance of selecting the appropriate CNN architecture for crash diagram analysis, particularly in the context of pedestrian safety. In addition to evaluating model performance, computational efficiency is also considered. This study contributes novel insights to pedestrian safety research by leveraging image processing technology and highlights CNNs potential in detecting concealed pedestrian crash patterns.

In the second part, a safety assessment framework is designed to identify and extract latent features from crash data, specifically focusing on understanding the factors influencing injury severity among vehicle and micro-mobility crashes in Michigan's urban areas. The information extraction technique is employed in this study to extract types of micro-mobility devices from crash narratives, providing insights into their usage patterns and contexts. Micro-mobility devices analyzed in this study are bicycles, e-wheelchairs, skateboards, and e-scooters. The AlexNet CNN architecture is utilized to enable the recognition and classification of micro-mobility device collision locations into three categories: roadside, shoulder, and bicycle lane. Subsequently, the Random Forest classification algorithm is utilized to pinpoint the primary factors and their interactions that affect the severity of micro-mobility injuries. Also, to tackle the challenge posed by the imbalanced dataset comprising fatal/severe injury crashes in contrast to those involving no/minor injuries, the Synthetic Minority Oversampling technique is implemented. This study highlights the importance of micro-mobility device type and crash location in determining injury severity and suggests several strategies to improve road safety in the context of micro-mobility.

The final framework integrates the AlexNet CNN architecture with Natural Language Processing (NLP) techniques to investigate injury outcomes of horse-and-buggy crashes in rural Michigan by mining pedestrian crash reports. The AlexNet model can recognize and identify horse-and-buggy crashes, while NLP techniques extract primary risk factors from crash narratives, including unigram and trigram analysis. Logistic regression analysis is used to predict severity levels based on significant trigrams. The findings highlight the critical risks associated with horse-and-buggy crashes and underscore the potential of advanced image-processing techniques in traffic safety research. This study suggests that integrating AI technologies into pedestrian safety studies could lead to more precise and actionable findings, enhancing road safety for all users.

Access Setting

Dissertation-Abstract Only

Restricted to Campus until

8-1-2034

This document is currently not available here.

Share

COinS