Enhanced Balanced-Generative Adversarial Networks to Predict Pedestrian Injury Types

Author:

Somvanshi Shriyank1,Antariksa Gian1,Das Subasish1

Affiliation:

1. Texas State University

Abstract

Abstract

Pedestrians are at a significantly higher risk of suffering serious injuries or death in road traffic incidents. In 2021 alone there were 7,388 pedestrian fatalities and 60,577 injuries. Addressing this critical issue, our study introduced a novel methodology for predicting the severity of pedestrian crashes. This method leveraged advanced deep learning models such as Inception-ResNet-v2, Inception-v3, and Xception applied to synthetic data generated by Generative Adversarial Networks (GANs). This study analyzed data related to pedestrian crash severity in Louisiana spanning five years (2016–2021), encompassing forty variables that include pedestrian demographics, accident location, and vehicle specifics. The severity of crashes was categorized into three categories: injury, no injury, and fatal. To combat the challenge of data imbalance, our study implemented a novel method that combined traditional sampling methods with GANs. This integrated methodology facilitated the generation of synthetic data utilizing the Conditional Tabular GAN (CTGAN) model and the attainment of balanced datasets by employing under-sampling via the Random Under Sampler (RUS) technique and over-sampling through the Synthetic Minority Over-sampling Technique (SMOTE). Thereafter, the DeepInsight technique was employed to transform numerical and categorical crash data into image format, making it compatible with the deep learning models utilized. The findings reveal that the models demonstrated improved predictive capabilities when applied to the over sampled dataset, which was achieved by increasing the number of instances in the minority class to balance the distribution of classes, as evidenced by various performance metrics including accuracy, precision, recall, and F1 score. Specifically, the Inception-ResNet-v2, Inception-v3, and Xception models recorded predictive accuracies of 82.73%, 84.75%, and 69.07% respectively, with the over sampled dataset, which was the highest among the three sampling categories of data considering all the metrics. The insights derived from this research have practical applications for urban planners, city engineers, safety professionals, transportation authorities, emergency service providers, vehicle manufacturers, and traffic management centers.

Publisher

Research Square Platform LLC

Reference44 articles.

1. NHTSA. Pedestrians Traffic Safety Facts 2021 Data. Publication 813458. U.S. Department of Transportation, Washington, D.C (2023) p. 16

2. WHO (2024) Road Traffic Injuries. https://www.who.int/news-room/fact-sheets/detail/road-traffic-injuries. Accessed Mar. 6

3. GHSA Pedestrian Traffic Fatalities by State: 2022 Preliminary Data | GHSA

4. Kendi S, Johnston BD, V.COUNCIL ON INJURY AND, POISON PREVENTION. Epidemiology and Prevention of Child Pedestrian Injury (2023) Pediatrics 152(1):e2023062508. https://doi.org/10.1542/peds.2023-062508

5. Real-Time Crash Risk Prediction on ArterialsBased on LSTM-CNN;Li P;Accid Anal Prev,2020

同舟云学术

1.学者识别学者识别

2.学术分析学术分析

3.人才评估人才评估

"同舟云学术"是以全球学者为主线,采集、加工和组织学术论文而形成的新型学术文献查询和分析系统,可以对全球学者进行文献检索和人才价值评估。用户可以通过关注某些学科领域的顶尖人物而持续追踪该领域的学科进展和研究前沿。经过近期的数据扩容,当前同舟云学术共收录了国内外主流学术期刊6万余种,收集的期刊论文及会议论文总量共计约1.5亿篇,并以每天添加12000余篇中外论文的速度递增。我们也可以为用户提供个性化、定制化的学者数据。欢迎来电咨询!咨询电话:010-8811{复制后删除}0370

www.globalauthorid.com

TOP

Copyright © 2019-2024 北京同舟云网络信息技术有限公司
京公网安备11010802033243号  京ICP备18003416号-3