The use of drones (UAV - Unmanned Aerial Vehicles) in environmental and infrastructural monitoring generates continuous streams of heterogeneous data, ranging from low-latency telemetry to heavy multimedia streams. Managing and processing this information imposes significant architectural challenges, highlighting the limitations of traditional centralized and polling systems in terms of latency, bandwidth consumption, and computational inefficiency. This thesis presents the design and development of an event-driven Cloud-Edge architecture based on Amazon Web Services (AWS), aimed at the advanced visual analysis of the monitored infrastructures and the optimization of cloud resources. The system decouples the UAV flight telemetry, managed in near real-time via AWS IoT Core and DynamoDB, from the acquisition of images routed to Amazon S3. The orchestration of computationally intensive workloads is entrusted to GPU-accelerated Amazon EC2 instances (g4dn.xlarge), within which a containerized Docker environment (managed via Amazon ECR) executes a cascaded inference using YOLO11n for spatial object detection and Qwen2-VL-2B for the semantic understanding of the scene. The innovative contribution of the project lies in the radical optimization of network traffic and operational costs. The architectural transition to Amazon SQS, leveraging the Long Polling logic, allowed to zeroed data traffic in the idle state, overcoming the inefficiencies of IoT protocols for purely cloud communication. Concurrently, an automatic de-provisioning system supervised by Amazon CloudWatch has been engineered to physically terminate the EC2 hardware by analyzing network traffic metrics, effectively minimizing economic waste. Historical archiving and asynchronous log analysis are finally guaranteed by the integration of Kinesis Firehose and Amazon Athena. The results obtained during the testing phase confirm the validity of the platform. The system has indeed met the operational constraints by processing video streams with latencies below 33.3 ms per frame (a critical parameter to ensure real-time processing at 30 FPS), efficiently managing the 16 GB VRAM physical limit on the NVIDIA Tesla T4 accelerator. From an economic perspective (FinOps), the auto-suspension logic generated a 91.8% reduction in Cloud Waste, shrinking the monthly infrastructure costs from $474.50 to $39. These metrics certify the realization of a scalable, resilient, and responsive platform capable of integrating the UAV as an intelligent peripheral node within an economically sustainable and computationally efficient Cloud ecosystem.
L’impiego di droni (UAV - Unmanned Aerial Vehicles) nel monitoraggio ambientale e infrastrutturale genera flussi continui di dati eterogenei, spaziando dalla telemetria a bassa latenza ai flussi multimediali pesanti. Gestire ed elaborare queste informazioni impone sfide architetturali significative, evidenziando i limiti dei tradizionali sistemi centralizzati e di polling in termini di latenza, consumo di banda e inefficienza computazionale. Questa tesi presenta la progettazione e lo sviluppo di un’architettura Cloud-Edge event-driven, basata sui servizi Amazon Web Services (AWS), finalizzata all’analisi visiva avanzata delle infrastrutture monitorate e all’ottimizzazione delle risorse cloud. Il sistema disaccoppia la telemetria di volo dell’UAV, gestita in tempo quasi reale tramite AWS IoT Core e DynamoDB, dall’acquisizione delle immagini, convogliate su Amazon S3. L’orchestrazione dei carichi di lavoro computazionalmente intensivi è affidata a istanze Amazon EC2 (g4dn.xlarge) dotate di accelerazione GPU, all’interno delle quali un ambiente containerizzato Docker (gestito tramite Amazon ECR) esegue un’inferenza a cascata utilizzando YOLO11n per il rilevamento spaziale degli oggetti e Qwen2-VL-2B per la comprensione semantica della scena. Il contributo innovativo del progetto risiede nella radicale ottimizzazione del traffico di rete e dei costi operativi. La transizione architetturale verso Amazon SQS, sfruttando la logica di Long Polling, ha permesso di azzerare il traffico dati in stato di inattività, superando le inefficienze dei protocolli IoT per la comunicazione puramente cloud. Parallelamente, è stato ingegnerizzato un sistema di de-provisioning automatico supervisionato da Amazon CloudWatch che interrompe fisicamente l’hardware EC2 analizzando le metriche del traffico di rete, minimizzando efficacemente gli sprechi economici. L’archiviazione storica e l’analisi asincrona dei log sono infine garantite dall’integrazione di Kinesis Firehose e Amazon Athena. I risultati ottenuti in fase di collaudo confermano la validità della piattaforma. Il sistema ha infatti rispettato i vincoli operativi elaborando i flussi video con latenze inferiori ai 33.3 ms per fotogramma (parametro critico per garantire l’elaborazione in tempo reale a 30 FPS), gestendo in modo efficiente il limite fisico dei 16 GB di VRAM sull’acceleratore NVIDIA Tesla T4. Sotto il profilo economico (FinOps), le logiche di autosospensione hanno generato un abbattimento del Cloud Waste pari al 91.8%, contraendo i costi mensili dell’infrastruttura da $474.50 a $39. Tali metriche certificano la realizzazione di una piattaforma scalabile, resiliente e reattiva, capace di integrare l’UAV come nodo periferico intelligente all’interno di un ecosistema Cloud economicamente sostenibile e computazionalmente efficiente.
Architettura Edge-to-Cloud per Sistemi UAV: Analisi Visiva Event-Driven tramite YOLO e Qwen su AWS
PEROZZI, MATTEO
2025/2026
Abstract
The use of drones (UAV - Unmanned Aerial Vehicles) in environmental and infrastructural monitoring generates continuous streams of heterogeneous data, ranging from low-latency telemetry to heavy multimedia streams. Managing and processing this information imposes significant architectural challenges, highlighting the limitations of traditional centralized and polling systems in terms of latency, bandwidth consumption, and computational inefficiency. This thesis presents the design and development of an event-driven Cloud-Edge architecture based on Amazon Web Services (AWS), aimed at the advanced visual analysis of the monitored infrastructures and the optimization of cloud resources. The system decouples the UAV flight telemetry, managed in near real-time via AWS IoT Core and DynamoDB, from the acquisition of images routed to Amazon S3. The orchestration of computationally intensive workloads is entrusted to GPU-accelerated Amazon EC2 instances (g4dn.xlarge), within which a containerized Docker environment (managed via Amazon ECR) executes a cascaded inference using YOLO11n for spatial object detection and Qwen2-VL-2B for the semantic understanding of the scene. The innovative contribution of the project lies in the radical optimization of network traffic and operational costs. The architectural transition to Amazon SQS, leveraging the Long Polling logic, allowed to zeroed data traffic in the idle state, overcoming the inefficiencies of IoT protocols for purely cloud communication. Concurrently, an automatic de-provisioning system supervised by Amazon CloudWatch has been engineered to physically terminate the EC2 hardware by analyzing network traffic metrics, effectively minimizing economic waste. Historical archiving and asynchronous log analysis are finally guaranteed by the integration of Kinesis Firehose and Amazon Athena. The results obtained during the testing phase confirm the validity of the platform. The system has indeed met the operational constraints by processing video streams with latencies below 33.3 ms per frame (a critical parameter to ensure real-time processing at 30 FPS), efficiently managing the 16 GB VRAM physical limit on the NVIDIA Tesla T4 accelerator. From an economic perspective (FinOps), the auto-suspension logic generated a 91.8% reduction in Cloud Waste, shrinking the monthly infrastructure costs from $474.50 to $39. These metrics certify the realization of a scalable, resilient, and responsive platform capable of integrating the UAV as an intelligent peripheral node within an economically sustainable and computationally efficient Cloud ecosystem.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Perozzi.pdf
accesso aperto
Dimensione
53.11 MB
Formato
Adobe PDF
|
53.11 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/29630