Nowadays, road safety represents one of the main challenges for urban mobility and requires increasingly advanced tools to analyze, predict, and interpret the factors that influence risk on the road network. This thesis aims to develop and extend the data-driven framework RoadSafeAI, a deep learning model based on a modified ResNet-18 architecture adapted to a regression task, whose objective is to estimate the road risk associated with each 200 m × 200 m cartographic tile generated from OpenStreetMap France. To validate the work carried out in the previous thesis, the robustness properties of the network were evaluated, including sensitivity to rotations and consistency with respect to translations, in order to analyze the model's ability to provide coherent estimates as the spatial configuration of the input varies. Further validation of the model capabilities is provided by the application of Explainable Artificial Intelligence (XAI) techniques, adapted to regression tasks, such as Gradient-weighted Class Activation Mapping (Grad-CAM), Layer-wise Relevance Propagation (LRP), and Guided Backpropagation. Through the analysis and visualization of these methods, the black-box nature of the model is mitigated, making it possible to identify the areas that contribute most to the risk estimate and to highlight the presence of possible patterns that may introduce bias into the predictions. Finally, the thesis proposes a traffic-aware version of the baseline model, called RoadSafeAI 2.0, in which traffic information is integrated with the road features learned by the network through the Feature-wise Linear Modulation (FiLM) mechanism. The results obtained through spatial cross-validation show that the model is able to adapt its predictions to different daily traffic conditions in the metropolitan area of Milan, providing a more dynamic and realistic representation of risk and offering a concrete contribution to the improvement of urban road safety.
Oggigiorno, la sicurezza stradale rappresenta una delle principali sfide per la mobilità urbana e richiede strumenti sempre più avanzati per analizzare, predire e interpretare i fattori che influenzano il rischio sulla rete stradale. Questa tesi si pone l'obiettivo di sviluppare ed estendere il framework data-driven RoadSafeAI, un modello di deep learning basato su un'architettura ResNet-18 modificata e adattata a un compito di regressione, il cui scopo è stimare il rischio stradale associato a ogni tile cartografica di 200 m × 200 m generata da OpenStreetMap France. Per validare il lavoro svolto nella precedente tesi, sono state valutate le proprietà di robustezza della rete, quali la sensibilità alle rotazioni e la coerenza rispetto alle traslazioni, in modo da analizzare la capacità del modello di fornire stime consistenti al variare della configurazione spaziale dell'input. Un'ulteriore validazione delle capacità del modello è fornita dall'applicazione di tecniche di Explainable Artificial Intelligence (XAI), adattate a compiti di regressione, quali Gradient-weighted Class Activation Mapping (Grad-CAM), Layer-wise Relevance Propagation (LRP) e Guided Backpropagation. Attraverso l'analisi e la visualizzazione di questi metodi, viene mitigata la natura black box del modello, consentendo di individuare le aree che maggiormente contribuiscono alla stima del rischio e di evidenziare la presenza di eventuali pattern che possono introdurre bias nelle predizioni. Infine, la tesi propone una versione traffic-aware del modello baseline, denominata RoadSafeAI 2.0, in cui l'informazione sul traffico viene integrata con le feature stradali apprese dalla rete mediante il meccanismo di Feature-wise Linear Modulation (FiLM). I risultati ottenuti mediante cross-validation spaziale mostrano che il modello è in grado di adattare le proprie predizioni alle diverse condizioni giornaliere di traffico nell'area metropolitana di Milano, restituendo una rappresentazione del rischio più dinamica e realistica e offrendo un contributo concreto al miglioramento della sicurezza urbana.
RoadSafeAI 2.0: Spatial and traffic aware deep learning framework for urban road risk prediction
Salina, Daniele
2025/2026
Abstract
Nowadays, road safety represents one of the main challenges for urban mobility and requires increasingly advanced tools to analyze, predict, and interpret the factors that influence risk on the road network. This thesis aims to develop and extend the data-driven framework RoadSafeAI, a deep learning model based on a modified ResNet-18 architecture adapted to a regression task, whose objective is to estimate the road risk associated with each 200 m × 200 m cartographic tile generated from OpenStreetMap France. To validate the work carried out in the previous thesis, the robustness properties of the network were evaluated, including sensitivity to rotations and consistency with respect to translations, in order to analyze the model's ability to provide coherent estimates as the spatial configuration of the input varies. Further validation of the model capabilities is provided by the application of Explainable Artificial Intelligence (XAI) techniques, adapted to regression tasks, such as Gradient-weighted Class Activation Mapping (Grad-CAM), Layer-wise Relevance Propagation (LRP), and Guided Backpropagation. Through the analysis and visualization of these methods, the black-box nature of the model is mitigated, making it possible to identify the areas that contribute most to the risk estimate and to highlight the presence of possible patterns that may introduce bias into the predictions. Finally, the thesis proposes a traffic-aware version of the baseline model, called RoadSafeAI 2.0, in which traffic information is integrated with the road features learned by the network through the Feature-wise Linear Modulation (FiLM) mechanism. The results obtained through spatial cross-validation show that the model is able to adapt its predictions to different daily traffic conditions in the metropolitan area of Milan, providing a more dynamic and realistic representation of risk and offering a concrete contribution to the improvement of urban road safety.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Salina_Tesi.pdf
non accessibile
Dimensione
59.42 MB
Formato
Adobe PDF
|
59.42 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Salina_Executive_Summary.pdf
non accessibile
Dimensione
2.73 MB
Formato
Adobe PDF
|
2.73 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260128