ISEL - Eng. Elect. Tel. Comp. - Dissertações de Mestrado
URI permanente para esta coleção:
Navegar
Entradas recentes
- Semantic segmentation of medical images for fast diagnosisPublication . Carvalho, António Maria Ferreira de Oliveira; Véstias, Mário PereiraAbstract Semantic segmentation of medical images is a critical task in modern healthcare, enabling precise identification and localization of anatomical structures and pathological regions by classifying image regions at the pixel level, thereby supporting rapid and accurate diagnosis. Although deep learning models have achieved remarkable performance in this domain, their deployment in real-world clinical environments is often constrained by the high computational demands of traditional software implementations, resulting in inefficiency and latency. Field-Programmable Gate Arrays (FPGAs) provide a promising alternative by offering programmable hardware-level acceleration tailored for deep learning inference tasks, with advantages such as parallel processing, low latency, and energy-efficient computation, making them particularly suitable for time-sensitive applications including tumor detection, organ delineation, and lesion classification. In this work, a lightweight neural network architecture, Mobile-CMUNeXt, was designed and optimized for FPGA deployment through architectural pruning, quantization-aware training, and hardware-friendly modifications, alongside the development of modular accelerator cores for depthwise, pointwise, and 3D convolutions implemented on the Avnet Ultra96-V2 platform. Experimental results demonstrate that the proposed accelerator achieves real-time inference performance with clinically acceptable segmentation accuracy, while balancing resource utilization and maintaining energy efficiency and low latency, and further show competitive throughput compared to CPU and GPU baselines, validating the effectiveness of the hardware–software co-design methodology. Overall, this thesis establishes that FPGA-based accelerators constitute a viable solution for deploying deep learning models in edge medical devices by combining efficient neural network design with custom hardware optimizations, thereby enabling real-time, privacy-preserving, and energy-efficient semantic segmentation of medical images.
- Guidance of autonomous vehicles through Visible Light CommunicationPublication . Carvalho, António Joaquim Martins de; Vieira, Maria Manuela de Almeida Carvalho; Louro, Paula Maria Garcia; Véstias, Mário PereiraAbstract Automated Guided Vehicles (AGVs) and Autonomous Mobile Robots (AMRs) are transforming industrial automation, delivering improvements in efficiency, safety, and operational accuracy. Despite these advancements, AGVs and AMRs still face limitations in indoor environments, where conventional navigation technologies may be unreliable or restricted. Visible Light Communication (VLC) emerges as a promising solution, exploiting the visible light spectrum to enable precise localization and reliable communication. By taking advantage of existing LED infrastructures, VLC provides accurate positioning and continuous real-time data exchange, offering a cost-effective and energy-efficient alternative to traditional systems. This dissertation examines the integration of VLC into AGV and AMR systems, with a focus on applications in structured environments such as factories, warehouses, and logistics hubs. The study also considers the main challenges associated with VLC, including line-of-sight dependence and susceptibility to ambient light interference, and discusses strategies to mitigate these issues. Within this framework, centralized coordination approaches are analyzed, highlighting how VLC can enhance navigation, task allocation, and traffic management in controlled environments. A simulation study based on the proposed VLC grid and a centralized coordinator was conducted in a grid-based warehouse environment. Results show that VLC delivers stable cell-level localization and supports coordinated operation. Across the evaluated scenarios, the rule-based baseline achieved the highest throughput with few collisions and no deadlocks, whereas a non-learning centralized controller was inefficient under load; preliminary learning-based control did not surpass the baseline within the tested horizons. Overall, the study confirms the feasibility of VLC-enabled centralized coordination and identifies practical constraints (line-of-sight, ambient light, channel load) together with clear avenues for improvement.
- Desenvolvimento e implementação de uma rede LoRa para monitorização sísmica com sensores acelerómetrosPublication . Azevedo, Rúben Miguel Paulo; Pires, Luís Miguel Rego; Fialho, Vítor Manuel de OliveiraA monitorização sísmica desempenha um papel central na mitigação dos riscos associados a fenómenos naturais. Contudo, os sistemas tradicionais de deteção apresentam custos elevados e cobertura geográfica limitada, o que dificulta a sua implementação em larga escala. Neste contexto, soluções baseadas em redes Internet of Things (IoT), aliadas a dispositivos e sensores de baixo custo e baixo consumo, surgem como uma alternativa promissora para o desenvolvimento de redes de monitorização mais acessíveis e flexíveis. Esta dissertação apresenta o desenvolvimento de um sistema de deteção sísmica de baixo custo e consumo energético, baseado em um microcontrolador e acelerómetro comercial. O sistema realiza a aquisição e processamento de dados de aceleração, aplica algoritmos para identificação de eventos sísmicos e transmite a informação através de uma rede Long Range/Long Range Wide Area Network (LoRa/LoRaWAN). Diversos testes foram realizados para a avaliar a cobertura da comunicação rádio, a fiabilidade da transmissão de dados e o consumo energético, de modo a demonstrar a viabilidade de sistemas baseados em LoRa/LoRaWAN como solução económica e eficaz para monitorização sísmica. Entre os vários resultados obtidos, destacam-se a cobertura de aproximadamente 3 km, para uma taxa de receção de pacotes igual ou superior a 90% em cenário suburbano a rural, e uma autonomia de cerca de 160 horas.
- Near surface object detection radar for UAVPublication . Pereira, Tiago Filipe Amado Ribeirete; Casaleiro, João Carlos Ferreira de Almeida; Costa, Vítor Manuel da SilvaAbstract The reliable detection of underground objects, such as landmines, remains a critical challenge in post-conflict areas where irregular soils complicate accurate target identification for demining, increasing risks to operators. This work addresses this issue by developing a ground-penetrating radar (GPR) system to be carried out by an unmanned aerial vehicle (UAV), using a Stepped Frequency Continuous Wave (SFCW) radar technique from 1 GHz to 6 GHz to transmit and to receive the signals, as well as self-interference cancelling techniques to cancel or mitigate self-interference and ground reflection signals. The GPR system was still able to function as a radar with the BladeRF 2.0 micro xA9, clearly detecting metal targets 1 meter away and distinguishing it from the floor and sand, although each data acquisition takes at least 10 seconds in order to have optimized values. The software cancellation was possible, but not through hardware.
- Applications performance analysis in C-ITSPublication . Sandu, Petru Lucian; Simão, José Manuel de Campos Lages Garcia; Serrador, António João NunesAbstract The rapid advancement of transport technology has enabled the deployment of Cooperative Intelligent Transport Systems (C-ITS), enhancing road safety, traffic flow, and overall efficiency. However, current implementations are designed for vehicles with dedicated On-Board Units (OBUs), limiting access for legacy vehicles. This presents a major barrier to widespread adoption and the full potential of C-ITS. This thesis evaluates the C-ITS architecture currently under development by Infraestruturas de Portugal (IP), focusing on its performance, scalability, and ability to support broader adoption. A mobile application was implemented to simulate the end-user experience, subscribing to C-ITS topics through MQTT and enabling practical assessment of message delivery to devices beyond dedicated OBUs. Comprehensive performance tests were conducted, including latency analysis, load testing, and resilience under component failures. This research contributes to the documentation and evaluation of a real-world C-ITS deployment, highlighting key areas for optimisation. Ultimately, it supports the ongoing efforts to scale C-ITS and ensure more inclusive access to traffic information for all vehicles.
- Implementation and evaluation of a relay system for LoRaWAN-based IoT networksPublication . Lobo, Leonor Sofia Rodrigues; Cota, Nuno António Fraga Juliano; Cruz, Nuno Miguel MachadoAbstract The rapid expansion of IoT has revolutionized various industries, yet its growth remains limited in remote areas where traditional network infrastructures, such as mobile networks and satellites, are either unreliable or too costly to implement. This thesis explores the use of LoRaWAN as a promising solution for low-power, long-range communication in remote locations. Despite its advantages, LoRaWAN faces challenges in areas with weak backhaul infrastructure or longdistance communication needs. To address these issues, this work proposes and implements a relay system composed of two devices: the C-Mesh, which acts as a gateway and listens continuously to end-devices messages and the C-Point, which encapsulates and forwards those uplinks. A dedicated communication protocol between C-Mesh and C-Point was developed, including a mechanism to update OTAA session keys over BLE. The potential of LoRa Mesh and satellite networks is also explored, along with the study of existing relay systems, such as the TS011-1.0.0 specification, that provide useful design insights. Field tests showed that the proposed relay increased delivery from 12.0% to 47.3% and raised the message receptions with SNR > 0 dB from 16.2% to 42.1%, confirming its role in extending LoRaWAN coverage and improving signal quality.
- Intelligent traffic intersection management through multi-agent reinforcement learningPublication . Antunes, Tomás Alexandre HenriquesAbstract Urban traffic management remains a persistent challenge for modern cities, particularly during peak hours when large volumes of vehicles and pedestrians converge, causing severe congestion, delays, and increased road safety risks. Given the limited feasibility of expanding physical infrastructure, it becomes essential to explore intelligent and adaptive solutions for traffic signal control. This work focuses on the application of Multi-Agent Reinforcement Learning (MARL) algorithms to optimize decision-making and coordination of traffic signals in urban networks. It is assumed that Visible Light Communication (VLC) between vehicles and infrastructure is available to provide real-time data required for the decision process, although this component is not the primary focus of the research. To validate the proposed approach, a simulation environment was developed using Simulation of Urban Mobility (SUMO), consisting of five interconnected signalized intersections. Within this context, different MARL algorithms were studied and compared, including Deep Q-Learning Network (DQN) and Multi-Agent Proximal Policy Optimization (MAPPO), with the objective of evaluating their performance under heterogeneous and dynamic traffic scenarios. The results show that MAPPO consistently outperforms DQN-based methods, achieving faster and more complete clearance of vehicles while maintaining lower waiting times for pedestrians. QT-DQN provides slight improvements over DQN in vehicle flow but at the cost of harming pedestrian performance. Overall, the study demonstrates that MARL methods, and particularly MAPPO, offer significant improvements in traffic efficiency and fairness, reinforcing their potential for deployment in real-world urban environments.
- Emotion recognition in multimedia contentPublication . Condesso, Sofia Fernandes; Ferreira, Artur Jorge; Leite, Nuno Miguel da Costa de SousaAbstract Emotion Recognition (ER) has become crucial in Human-Computer Interaction (HCI), with applications ranging from mental health support to adaptive learning. While many existing approaches rely on controlled environments or hardware-based sensors, this thesis explores non-contact unimodal methods—speech, facial expressions, and textual data—for a more naturalistic and practical analysis of emotions. First, we conduct a systematic evaluation of unimodal ER, comparing classical Machine Learning (ML) and Deep Learning (DL) approaches across multiple unimodal and multimodal datasets. For speech modality (audio), we extract acoustic features using openSMILE (GeMAPS), and learn with models such as Support Vector Machines (SVM) and Random Forests. Results show that feature selection on acoustic features can improve Speech Emotion Recognition (SER). For Facial Emotion Recognition (FER), we experiment with DeepFace and a lightweight Convolutional Neural Networks (CNN). For textual emotion recognition, we employ Word2Vec and GloVe with ML and DL models, and also experiment zero-shot and few-shot learning with large language models. In multimodal experiments, fusion of text and audio modalities improved accuracy to 0.45, confirming the benefit of combining complementary emotional cues. However, adding the visual modality led to a slight degradation in performance, attributed to suboptimal frame sampling. Overall, results highlight the trade-offs between unimodal simplicity and multimodal robustness, demonstrating that lightweight, interpretable models can achieve practical performance for real-world emotion-aware applications.
- Standard-based smart card access control architecture for critical infrastructuresPublication . Oliveira, António Filipe Ramic Santos; Dias, Tiago Miguel Braga da Silva; Chaves, Ricardo Jorge FernandesAbstract This thesis proposes an enterprise-centered access control architecture that is technically compatible with widely adopted standards, enabling the convergence of physical and logical access on a single smart card credential while preserving interoperability and enterprise control over identity proofing, registration, and credential issuance. This need is particularly acute for operators of critical infrastructure, who must replace legacy badge-based systems, ensure cross-site and outage resilience, unify access control with support for rapid revocation, generate audit trails, and avoid vendor lock-in through the use of open standards. The proposed architecture was built to meet these needs. The main outcomes of this work are a standards-aligned enterprise architecture for a multisite access control system and a smart card credential with a well-defined, standardized data model and lifecycle. Additionally, this work showcases a proof-of-concept prototype consisting of a Card Management System that personalizes cards and issues Public Key Infrastructure (PKI) credentials. The prototype also features a Java Card applet that implements a standardized data model and command set, supporting asymmetric cryptography-based multi-factor authentication mechanisms. Furthermore, it includes an access control system emulator for testing purposes. Using this stack, it is possible to issue cards and test them by performing strong authentication mechanisms, demonstrating the end-to-end feasibility of the devised solution for both physical and logical access scenarios.
- Aplicação para caracterização da vozPublication . Santos, Pedro Branco; Cordeiro, Hugo TitoEste projeto propõe uma aplicação simples e intuitiva com a capacidade de se ligar a um servidor, com o foco a melhorar a experiência do utilizador através de dinamismo e modelação. A funcionalidade principal da aplicação é apresentar resultados, numa interface, que caracterizam a fala através da comunicação com o servidor, mantendo sempre a troca de informação segura e que não haja perdas desta. Um dos objetivos principais do projeto foi garantir que a experiência do utilizador seja agradável, removendo qualquer complexidade na interface da aplicação. A arquitetura do servidor está desenhada de modo a garantir que a informação é sempre processada e que o utilizador obtenha sempre os resultados desejados. Este projeto desenvolveu uma aplicação protótipo que permite a extração de parâmetros dos sinais de fala para rastreio de patologias da voz. A aplicação pretende ser dinâmica, nomeadamente através da inclusão de novos algoritmos. A interface de utilizador foi desenvolvida em Unity3D de modo a permitir o uso em múltiplas plataformas. Os algoritmos foram desenvolvidos em Python de modo a tirar partido das bibliotecas de processamento de fala e sinal existentes.