Scarano, Antonella (2024) Innovative approaches for safety analysis of cyclists. [Tesi di dottorato]

	Documento PDF Scarano_Antonella_37_COMPLETO.pdf Visibile a [TBR] Amministratori dell'archivio Download (10MB) | Richiedi una copia
	Documento PDF Scarano_Antonella_37_PARZIALE.pdf Visibile a [TBR] Amministratori dell'archivio Download (9MB) | Richiedi una copia

Abstract

Governments worldwide are increasingly investing in promoting sustainable transportation modes, such as cycling, to encourage greener cities. However, despite these efforts, cyclist safety remains a significant concern. Furthermore, greater risk can discourage road users from cycling (Osama and Sayed, 2017). Global statistics highlight the road safety issue, with over 1.19 million people annually killed in road traffic crashes (WHO, 2023). Vulnerable road users (VRUs), including pedestrians, cyclists, and motorcyclists, represent over half of these casualties, resulting in substantial economic burdens due to treatment costs and productivity loss. Cyclist safety has long been a critical area in road safety research and has recently gained increasing attention. Improving cyclist safety presents a different challenge compared to motorized vehicles, which have been extensively studied in literature. A starting point for expanding the knowledge base on cyclist safety is to carry out a systematic literature analysis to highlight research directions as well as existing research gaps. For this purpose, Bibliometrix-R tool was used to analyse 1066 documents retrieved from Web of Science (WoS) between 2012 and 2021. However, without understanding the factors contributing to the cyclist crash, the ability to implement appropriate and context-specific measures is severely limited. Thus, this doctoral dissertation developed and implemented both statistical models and Machine Learning (ML) techniques to analyse cyclist safety issues using crash data from Great Britain from 2016 to 2019. The statistical models employed include the Random Parameters Logit model and Random Parameters Logit with heterogeneity in means and variances. In contrast, the ML models encompass supervised approaches, such as Classification And Regression Tree (CART), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) explained by the SHapley Additive exPlanations (SHAP), as well as unsupervised methods, including cluster analysis. Three different RF algorithms were performed such as the traditional RF, the Weighted Subspace RF, and the Random Survival Forest. The latter demonstrated higher prediction accuracy. Furthermore, the research retraced each path from the root node to the leaf nodes of all the trees in the forest to generate if-then rules, selecting the most relevant rules based on their frequency of appearance in the trees. Each modelling approach also contributes unique insights to the overall understanding. Statistical models offer quantitative, easily interpretable results and are particularly effective in capturing heterogeneity in both means and variances. Machine Learning models, on the other hand, excel in automatically detecting patterns and relationships within the data, revealing hidden correlations and interactions between various factors that statistical models may not capture. A significant contribution of this research is the introduction of a novel hybrid approach for analysing factors influencing the severity of cyclist-related crashes by combining eXtreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) and a Random Parameters Logit model with heterogeneity in means and variances. This study is pioneering in applying such a hybrid method to analyse cyclist crash severity, filling a gap in the current literature. This model aims to address the limitations inherent in both ML and statistical models, providing a more robust analytical framework. The XGBoost-SHAP model reduced data dimensionality allowing the application of a robust statistical model, while the Random Parameters Logit model with heterogeneity in means and variances captured unobserved heterogeneity in both means and variances (RPLHMV). The statistical model identified 10 significant variables with fixed parameters for the fatal crashes, 22 significant variables for the serious injuries, and two indicator variables such as cyclist age ≤17 and the second vehicle manoeuvre equal to overtaking with statistically significant random parameters associated with serious injury outcomes. The relationships revealed by the logit framework were further examined using the XGBoost-SHAP, which provided deeper insights into the interactions between random and fixed parameters. Moreover, cluster analysis was utilized to deepen the understanding of cyclist safety issues by identifying homogeneous groups of cyclist crashes with similar characteristics. The k modes clustering algorithm was applied identifying 6 distinct typical crash scenarios. To further characterize these scenarios, a Binomial test was conducted to identify statistically significant modes within each cluster. Additionally, XGBoost - SHAP was applied to each scenario to determine which variables influence the severity of cyclist crashes, offering deeper insights into the factors involved. This dual approach, identifying typical crash scenarios and evaluating the key characteristics influencing crash severity within each scenario, provides a more understanding of the factors that contribute to cyclist safety issue. Based on the identified contributory factors, safety countermeasures useful to develop strategies for making bike a safer and more friendly form of transport were recommended.

Downloads

Actions (login required)

Modifica documento