This thesis addresses the topic of fire safety in hospital environments, with particular focus on early fire detection systems and decision support during emergency situations. Healthcare facilities present distinctive characteristics — a high proportion of non autonomous individuals and complex evacuation procedures — that make it necessary to adopt advanced tools capable of improving the speed and effectiveness of emergency response. The work begins with an analysis of the regulatory framework and data related to fire incidents in healthcare structures, and then develops a system based on deep learning models from the YOLO family, trained for the automatic recognition of smoke and flames from images and video streams captured by surveillance cameras. The experimental phase includes training on heterogeneous datasets covering both outdoor and indoor environments, followed by validation on real video footage. Results show a significant improvement in performance as dataset size and variety increase, while also highlighting certain limitations, particularly the presence of false positives caused by the visual similarity between flames and artificial light sources. An original contribution of the thesis concerns the integration of the fire detection system with information related to patient evacuation priorities. Given the impossibility of directly inferring the condition of a bedridden patient from the image alone, an approach based on visual markers associated with patients is proposed, which can be detected by the model. The system is structured as a multifactorial architecture in which the visual detection of smoke, flames and markers is combined with a decision-making logic to automatically identify the most critical areas and support emergency management. The work concludes with the proposal of an integrated model for hospital fire safety that combines computer vision, data analysis and decision logic, outlining its potential, limitations and possible future developments, including integration with IoT technologies and intelligent emergency management platforms.
La presente tesi affronta il tema della sicurezza antincendio in ambito ospedaliero, con particolare attenzione ai sistemi di rilevazione precoce degli incendi e al supporto decisionale in fase di emergenza. Gli ambienti sanitari presentano caratteristiche peculiari, elevata presenza di persone non autonome e complessità delle procedure di evacuazione che rendono necessario l'utilizzo di strumenti avanzati in grado di migliorare rapidità ed efficacia degli interventi. Il seguente lavoro parte da un'analisi del contesto normativo e dei dati relativi agli incendi in strutture sanitarie, per poi sviluppare un sistema basato su modelli di deep learning della famiglia YOLO, addestrati per il riconoscimento automatico di fumo e fiamme a partire da immagini e flussi video provenienti da telecamere di sorveglianza. La fase sperimentale comprende l'addestramento su dataset eterogenei, sia in ambienti outdoor che indoor, e la validazione su video reali. I risultati evidenziano un miglioramento significativo delle prestazioni all'aumentare della dimensione e varietà del dataset, pur rilevando alcune criticità legate alla presenza di falsi positivi dovuti alla somiglianza visiva tra fiamme e sorgenti luminose artificiali. Un contributo originale della tesi riguarda l'integrazione del sistema di rilevazione incendi con informazioni relative alla priorità evacuativa dei pazienti. Considerata l'impossibilità di dedurre direttamente la condizione di paziente allettato dalla sola immagine, viene proposto un approccio basato su marker visivi associati ai pazienti, rilevabili dal modello. Il sistema è strutturato come un'architettura multifattoriale in cui il rilevamento visivo di fumo, fiamme e marker viene combinato con una logica decisionale per individuare automaticamente le aree a maggiore criticità e supportare la gestione dell'emergenza. Il lavoro si conclude con la proposta di un modello integrato per la sicurezza antincendio ospedaliera che combina visione artificiale, analisi dei dati e logiche decisionali, delineando potenzialità, limiti e possibili sviluppi futuri, tra cui l'integrazione con tecnologie IoT e piattaforme intelligenti di gestione delle emergenze.
Sviluppo di un sistema intelligente basato su YOLO per la rilevazione integrata di incendi e la gestione delle priorità di evacuazione in ambiente ospedaliero
Fazio, Mattia
2025/2026
Abstract
This thesis addresses the topic of fire safety in hospital environments, with particular focus on early fire detection systems and decision support during emergency situations. Healthcare facilities present distinctive characteristics — a high proportion of non autonomous individuals and complex evacuation procedures — that make it necessary to adopt advanced tools capable of improving the speed and effectiveness of emergency response. The work begins with an analysis of the regulatory framework and data related to fire incidents in healthcare structures, and then develops a system based on deep learning models from the YOLO family, trained for the automatic recognition of smoke and flames from images and video streams captured by surveillance cameras. The experimental phase includes training on heterogeneous datasets covering both outdoor and indoor environments, followed by validation on real video footage. Results show a significant improvement in performance as dataset size and variety increase, while also highlighting certain limitations, particularly the presence of false positives caused by the visual similarity between flames and artificial light sources. An original contribution of the thesis concerns the integration of the fire detection system with information related to patient evacuation priorities. Given the impossibility of directly inferring the condition of a bedridden patient from the image alone, an approach based on visual markers associated with patients is proposed, which can be detected by the model. The system is structured as a multifactorial architecture in which the visual detection of smoke, flames and markers is combined with a decision-making logic to automatically identify the most critical areas and support emergency management. The work concludes with the proposal of an integrated model for hospital fire safety that combines computer vision, data analysis and decision logic, outlining its potential, limitations and possible future developments, including integration with IoT technologies and intelligent emergency management platforms.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Fazio.pdf
accessibile in internet per tutti
Descrizione: testo tesi
Dimensione
4.5 MB
Formato
Adobe PDF
|
4.5 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/10589/261470