Loading...
Publication
Estudo da previsão da gravidade dos acidentes rodoviários e dos fatores contribuintes utilizando machine learning
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.17 MB | Adobe PDF |
Advisor(s)
Abstract(s)
Os acidentes rodoviários representam uma das principais causas de morte e ferimentos
em Portugal. Em 2023, mais da metade dos acidentes com vítimas (52,7%) foi
provocada por colisões, os despistes representaram 33,6% dos casos, enquanto os
atropelamentos corresponderam a 13,4%, resultando em 479 mortes, 2.646 feridos
graves e 42.890 feridos leves. Diante desse cenário, este estudo tem como objetivo
desenvolver modelos preditivos para a gravidade dos acidentes rodoviários do tipo
colisão e despiste, utilizando técnicas de aprendizado de máquina (Machine Learning),
com base em dados fornecidos pela Autoridade Nacional de Segurança Rodoviária
(ANSR) referentes ao período de 2019 a 2023.
Para tratar o desequilíbrio entre as classes de gravidade da variável de resposta (feridos
leves vs. feridos graves e vítimas fatais), foram aplicadas as técnicas de balanceamento
SMOTE (Synthetic Minority Over-sampling Technique) e RUS (Random UnderSampling). Os algoritmos analisados foram: Naive Bayes (NB), Artificial Neural
Network (ANN), Support Vector Machine (SVM), Decision Tree (DT), Random Forest
(RF) e K-Nearest Neighbors (KNN). Os modelos foram treinados utilizando a validação
cruzada com 10 subconjuntos e a divisão percentual 80/20 (80% para treinamento e
20% para teste).
Os resultados demonstraram que os melhores desempenhos foram obtidos com o
algoritmo Random Forest, especialmente quando aplicado à base de dados balanceada
com a técnica SMOTE, apresentando para os acidentes do tipo colisão uma área sob a
curva ROC de 0,886 e um coeficiente Kappa igual a 0,607, já para os acidentes do tipo
despiste a área sob a curva ROC apresentou um valor de 0,871 e um coeficiente Kappa
de 0,573. A análise dos modelos também permitiu a identificação das variáveis mais
influentes nos acidentes (fatores de risco), possibilitando a definição de recomendações
para intervenções na infraestrutura viária e comportamento dos condutores, com vista
à redução dos acidentes com consequências mais graves. Permitiu ainda a realização
de estudos de previsão da gravidade dos acidentes mediante a modificação desses
fatores na base de dados original. Os resultados desta pesquisa evidenciam o potencial
do uso de técnicas de Machine Learning como ferramenta de apoio à tomada de
decisão na área da segurança viária.
Road accidents are one of the main causes of death and injury in Portugal. In 2023, more than half of accidents with victims (52.7%) were caused by collisions, run-off-road accidents represented 33.6% of cases, while run-overs accounted for 13.4%, resulting in 479 deaths, 2,646 serious injuries and 42,890 minor injuries. In view of this scenario, this study aims to develop predictive models for the severity of collision and run-offroad accidents, using machine learning techniques, based on data provided by the Portuguese Road Safety Authority (ANSR) for the period from 2019 to 2023. To address the imbalance between the severity classes of the response variable (minor injuries vs. serious injuries and fatalities), the SMOTE (Synthetic Minority Oversampling Technique) and RUS (Random Under-Sampling) balancing techniques were applied. The algorithms analysed were: Naive Bayes (NB), Artificial Neural Network (ANN), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF) and K-Nearest Neighbors (KNN). The models were trained using 10-fold cross-validation and a 80/20 percentage split (80% for training and 20% for testing). The results showed that the best performances were obtained with the Random Forest algorithm, especially when applied to the balanced database with the SMOTE technique, presenting for collision-type accidents an area under the ROC curve of 0.886 and a Kappa coefficient of 0.607, while for run-off-road accidents the results were an area under the ROC curve of 0.871 and a Kappa coefficient of 0.573. The analysis of the models also allowed the identification of the most influential variables in accidents (risk factors), enabling the definition of recommendations for interventions in road infrastructure and driver behavior, with a view to reducing accidents with more serious consequences. It also enabled studies to predict accident severity by modifying these factors in the original database. The results of this research highlight the potential of using Machine Learning techniques as a tool to support decision-making in the area of road safety.
Road accidents are one of the main causes of death and injury in Portugal. In 2023, more than half of accidents with victims (52.7%) were caused by collisions, run-off-road accidents represented 33.6% of cases, while run-overs accounted for 13.4%, resulting in 479 deaths, 2,646 serious injuries and 42,890 minor injuries. In view of this scenario, this study aims to develop predictive models for the severity of collision and run-offroad accidents, using machine learning techniques, based on data provided by the Portuguese Road Safety Authority (ANSR) for the period from 2019 to 2023. To address the imbalance between the severity classes of the response variable (minor injuries vs. serious injuries and fatalities), the SMOTE (Synthetic Minority Oversampling Technique) and RUS (Random Under-Sampling) balancing techniques were applied. The algorithms analysed were: Naive Bayes (NB), Artificial Neural Network (ANN), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF) and K-Nearest Neighbors (KNN). The models were trained using 10-fold cross-validation and a 80/20 percentage split (80% for training and 20% for testing). The results showed that the best performances were obtained with the Random Forest algorithm, especially when applied to the balanced database with the SMOTE technique, presenting for collision-type accidents an area under the ROC curve of 0.886 and a Kappa coefficient of 0.607, while for run-off-road accidents the results were an area under the ROC curve of 0.871 and a Kappa coefficient of 0.573. The analysis of the models also allowed the identification of the most influential variables in accidents (risk factors), enabling the definition of recommendations for interventions in road infrastructure and driver behavior, with a view to reducing accidents with more serious consequences. It also enabled studies to predict accident severity by modifying these factors in the original database. The results of this research highlight the potential of using Machine Learning techniques as a tool to support decision-making in the area of road safety.
Description
Keywords
Acidentes Rodoviários Fatores de Risco Gravidade dos Acidentes Machine
Learning Previsão Segurança Rodoviária Simulação