Ciência da Computação
URI permanente para esta coleçãohttps://locus.ufv.br/handle/123456789/197
Navegar
254 resultados
Resultados da Pesquisa
Item Estratégias de acolhimento versus evasão discente no primeiro ano: uma abordagem envolvendo cursos superiores na área de computação da Universidade Federal de Viçosa(Universidade Federal de Viçosa, 2025-12-05) Mundim, Pedro Cardoso de Carvalho; Silva, Thais Regina de Moura Braga; http://lattes.cnpq.br/0605148561935153A evasão no ensino superior é um dos principais desafios enfrentados pelas instituições, com impactos acadêmicos, sociais e financeiros significativos. Entre as áreas de conhecimento, os cursos da área de computação merecem destaque por apresentarem elevadas taxas de evasão, em grande parte devido à dificuldade nos primeiros períodos, quando disciplinas introdutórias como Programação são ofertadas. Nos últimos anos, observa-se uma crescente utilização de técnicas de análise exploratória e de aprendizado de máquina para identificar padrões e prever a evasão, geralmente com base em atributos de dados pessoais, desempenho acadêmico e quantidade de reprovações. No entanto, esses atributos dependem de informações obtidas após certo tempo de curso, o que limita sua utilidade em ações preventivas. Dessa forma, torna-se relevante investigar fatores disponíveis desde o ingresso do estudante, como as estratégias de acolhimento, que surgem como alternativas promissoras, especialmente no primeiro ano do curso, mas ainda pouco exploradas na literatura. Assim, o objetivo desta pesquisa é avaliar se as estratégias de acolhimento adotadas nos cursos de computação da Universidade Federal de Viçosa (UFV) possuem impacto na evasão no primeiro ano. Os resultados indicaram que a adoção progressiva dessas estratégias esteve associada a reduções de até 20 pontos percentuais nas taxas médias de evasão. Além disso, a integração dos dados acadêmicos e sociodemográficos com os atributos sobre acolhimento elevou o desempenho de modelos de aprendizado de máquina, com destaque para a Regressão Logística e o Support Vector Machine (SVM). Esses resultados indicam que tais atributos podem funcionar como indicadores precoces de evasão e, ao mesmo tempo, sugerem uma possível relação com o fortalecimento do sentimento de pertencimento dos calouros, fornecendo suporte para intervenções mais direcionadas por parte das instituições de ensino. Palavras-chave: evasão no ensino superior; aprendizado de máquina; estratégias de acolhimento; cursos de computação; evasão no primeiro anoItem Computer vision methods for aerial and ground-based farm monitoring: addressing the labeled data scarcity problem(Universidade Federal de Viçosa, 2025-11-24) Ferreira, Juliana Quintiliano de Oliveira; Silva, Michel Melo da; http://lattes.cnpq.br/3028081422884505The monitoring of agricultural areas is essential to ensure a safe environment, avoid economic losses, and prevent risks to infrastructure and human safety. Effective monitoring can detect lost animals, unauthorized human access, wild animal intrusions, among other issues. Artificial Intelligence (AI) is a powerful tool to automate this process through the processing of images captured from ground or aerial perspectives. However, supervised models require large volumes of labeled data to achieve good performance, and although data exists, most of it is private and inaccessible to the public or available in limited quantity. The available datasets are generally unlabeled or do not cover domain-specific scenarios, such as rural and farm environments. In addition, data collection in agricultural settings presents further challenges, such as the need for drones or other remote sensing technologies, making the process more expensive and complex. Finally, the annotation stage is also a bottleneck, as beyond collecting images, intensive manual work is required to label them, increasing both cost and time. In this context, we propose to address the problem of data scarcity and annotation complexity from two perspectives, aiming to mitigate the bottleneck of training AI models when only a small amount of labeled data is available. The first study focuses on aerial monitoring using images collected by Unmanned Aerial Vehicles (UAVs) on farms for the task of semantic segmentation. Semantic segmentation brings significant benefits to agricultural monitoring by automatically identifying and differentiating important elements of the rural environment, such as vegetation areas, bodies of water, and buildings. By precisely mapping these elements, the technique enables the identification of risk situations for livestock and infrastructure, contributing to safer and more efficient farm management. Thus, we investigated pre-training strategies using synthetic data from the same domain and real data from slightly different domains. We then fine-tuned on the target dataset, and the quantitative and qualitative results demonstrated that pre- training with the synthetic dataset achieved better final performance, leading to an increase of 3.1 p.p. in IoU, 6.4 in F1-Score, and 7.5 in Recall compared to the cross- domain real-image pre-training strategy. In the second study, we focus on object detection using ground-level images similar to security camera footage. This task is important in agricultural monitoring because it allows automatic identification and localization of animals and people, supporting security, livestock management, and the tracking of activities on the farm. To address the data scarcity challenge in this scenario, we proposed a method to effectively use multiple datasets even when they do not share the same classes, ensuring comprehensive coverage of all required categories. The proposed SmartClass methodology achieved more robust and adaptable detection approaches suitable for agricultural environments, with significant increases in Recall, mAP50, and mAP50-95 metrics compared to models trained without the methodology, thus demonstrating improved efficiency and reliability of the model. Keywords: artificial intelligence; computer vision; farm monitoring.Item A machine learning approach for predicting pallet demands in ceramic tile production(Universidade Federal de Viçosa, 2025-09-15) Oliveira, Matheus Aguilar de; Santos, André Gustavo dos; http://lattes.cnpq.br/3716978487689658This dissertation develops and evaluates a predictive approach for a variant of the Distributor’s Pallet Loading Problem (DPLP) applied in the context of an Italian ceramic tile manufacturer. In this scenario, orders from clients are composed by boxes of different sizes and weights that must be loaded onto pallets while meeting constraints on weight, volume and stability, as well as operational uncertainties in the warehouse. The objective is to estimate quickly and accurately the total number of pallets, and in an extended setting, the number of pallets of each type, required to fulfill an order. The proposed solution is a hybrid method that combines machine learning (ML) with heuristics. Historical data of the company is used to extract relevant features, which are then enriched with heuristic bounds to improve model accuracy. Three ML models, namely XGBoost, LightGBM, and Random Forest were trained and tuned using company's data. Experiments on thousands of past orders and a separate set of recent, unseen orders show that the hybrid approach consistently outperforms PackVol (the company's current software), achieving mean squared and absolute errors around 5.3 and 2.3 times smaller, respectively, while producing predictions a much faster time. The approach is also extended to a multi- output regression setting to predict pallet quantities by type, maintaining high accuracy and efficiency even under this more complex objective, with squared and absolute errors around 3.4 and 1.8 times smaller, respectivelly. The results demonstrate that combining ML with heuristic features yields a practical, scalable, and accurate predictive tool for the company's scenario, with potential applicability to other packing, loading, and logistics problems in this industry and many others. Keywords: distributor’s pallet loading problem; machine learning; heuristics ; tile production; real-world instancesItem ArchChain: proposta de um token de cessão para avaliação de custo e desempenho em redes blockchain(Universidade Federal de Viçosa, 2025-09-19) Mendonça, Ronan Dutra; Nacif, José Augusto Miranda; http://lattes.cnpq.br/2158451110718949Em meio ao crescente uso de Blockchain em aplicações públicas e privadas, uma lacuna persiste na análise sistemática da infraestrutura necessária para a participação eficiente em tais redes, sobretudo no que diz respeito aos nós de acesso a rede. O objetivo é criar um método que permita medir o impacto do uso de endpoints próprios em redes Blockchain. Este trabalho apresenta uma abordagem com a proposta de uma metodologia de avaliação de custo e desempenho para infraestrutura de redes Blockchain. A metodologia criada permite identificar o melhor compromisso entre custo e desempenho computacional na operação de um nó Blockchain em resposta a operações realizadas por uma aplicação descentralizada (Dapp). A arquitetura experimental considerou tanto redes públicas quanto permissionadas, com expansão entre os dois modelos. Os resultados indicaram que infraestruturas computacionais de menor porte podem atender satisfatoriamente a requisitos de desempenho em redes públicas, desde que bem escalonadas. Outro ponto relevante foi a introdução de um modelo para cálculo do compromisso entre custo e desempenho, fornecendo parâmetros concretos para a escolha de configurações ótimas. Os experimentos demonstraram que é viável manter a estabilidade de custo por transação em redes públicas através do ajuste da capacidade computacional. Também foi avaliada a interoperabilidade entre Blockchain, destacando o compromisso em tempo e custo, a partir da aplicação do Token de Cessão e da comparação entre protocolos de comunicação \textit{cross- chain}. Entre as principais contribuições da pesquisa, destacam-se: (i) a formulação de uma metodologia de avaliação de nós Blockchain sob as métricas de custo e desempenho; (ii) a criação de um mecanismo padronizado de cessão temporária de propriedade via tokens; e (iii) a expansão da funcionalidade desses tokens para além de uma única rede, por meio da interoperabilidade. Tais avanços contribuem com subsídios para escolha e adoção de soluções blockchain com maior eficiência e previsibilidade de custos. Palavras-chave: blockchain; interoperabilidade; cross-chain; token de cessão; avaliação de custo e desempenhoItem Vision-based gesture classifier for UAV teleoperation(Universidade Federal de Viçosa, 2025-02-28) Alves, Wérikson Frederiko de Oliveira; Brandão, Alexandre Santos; http://lattes.cnpq.br/8501121544866799Gesture-based control emerges as a promising alternative for the teleoperation of unmanned aerial vehicles (UAVs), offering an intuitive and natural interaction paradigm. Traditional control methods, such as joysticks and pre-programmed commands, often present usability challenges, particularly in dynamic outdoor environments where rapid decision-making and hands-free operation are essential. However, real-world implementation of gesture-based UAV control systems remains challenging due to environmental variability, lighting conditions, and user-specific differences in gesture execution. This dissertation designs and implements a UAV- integrated gesture recognition system optimized for outdoor teleoperation. The proposed approach leverages computer vision and machine learning techniques to enable robust, real-time gesture classification without reliance on specialized infrastructure. Initially, a motion capture-based approach (OptiTrack) facilitates experiments in controlled environments. To enhance adaptability for outdoor settings, the system transitions to a vision-based model utilizing YOLOv8, MediaPipe Hands, and BlazePose for hand and body tracking, combined with K-Nearest Neighbors (KNN) classification. Additionally, an onboard servo-visual control module, implemented using an ESP32CAM, enables the UAV to dynamically adjust its camera orientation and maintain continuous user tracking. Experimental validation demonstrates that the proposed system achieves a gesture classification accuracy of 92.57% across 14 predefined gestures, maintaining real-time performance even in outdoor conditions. The research also explores gesture-based UAV applications in light painting, showcasing the system's versatility for both creative and operational use cases. Despite these advancements, challenges such as gesture variability, low- light conditions, and image transmission latency persist. By addressing key limitations in real-world gesture recognition, this dissertation advances human-robot interaction (HRI) and UAV teleoperation, providing a scalable and adaptable framework for autonomous aerial systems controlled via gestures. Keywords: Gesture Recognition; UAV Teleoperation; Computer Vision; Human-Robot Interaction; Real-Time ClassificationItem Reinforcement learning applied to robot navigation(Universidade Federal de Viçosa, 2025-03-28) Carvalho, Kevin Braathen de; Brandão, Alexandre Santos; http://lattes.cnpq.br/1918730771175641Path planning is a key aspect of autonomous navigation, especially for autonomous vehicles, where different priorities such as path length, safety, and energy consumption must be considered. Traditional approaches, including dynamic programming and geometric methods, have been widely used to tackle this problem. However, in recent years, artificial intelligence techniques, particularly reinforcement learning, have gained increasing attention. This thesis explores the application of reinforcement learning methods, such as Q-learning, combined with other machine learning techniques like transfer learning, to improve convergence speed and overall performance in various robotic navigation tasks. The research begins with the development of offline global path planning algorithms for both 2D and 3D environments, validated through simulations and real-world experiments. The approach is then adapted for dynamic scenarios, enabling real-time local path planning, ultimately leading to six published papers. The proposed global path planning algorithms can flexibly balance three key priorities while ensuring that the agent can reach its destination from any starting point in the map. This provides robustness against external and internal disturbances. The local path planning adaptation maintains these priorities while operating in real time. Additional contributions include applications in intelligent logistics, such as automated warehouse organization. A curriculum-based training approach was introduced, progressively increasing task difficulty to facilitate learning. This was validated through simulations, with the results currently being finalized in writing. Finally, the thesis discusses the impact of state representation using depth sensors in mapless navigation. It also examines how different hyperparameter settings and Q-table initializations affect the performance of the proposed global path planning algorithms. Keywords: Robotic Navigation; Reinforcement Learning; Transfer Learning; Path PlanningItem Robotics studies with PINEL: a Platform for Interactive Navigation in Education and Learning contexts(Universidade Federal de Viçosa, 2025-10-02) Pinel, Guilherme Serra Francisco; Brandão, Alexandre Santos; http://lattes.cnpq.br/6341701225784770The rapid evolution of technology has driven changes in education, demanding new strategies for teaching programming, robotics, and computational thinking. This dissertation presents PINEL (Platform for Interactive Navigation in Education and Learning), a block-based programming platform integrated with the Robot Operating System (ROS 2) and a three-dimensional simulation environment. Designed as a modular, accessible, and extensible solution, PINEL bridges high-level visual abstractions and low-level control through an intermediate JSON representation, connecting a Blockly frontend, a Python/Flask backend with ROS 2 and closed-loop control, and a simulation connector based on the AuRoRA/MATLAB framework. Validation proceeded progressively: (i) in simulation, through experiments involving geometric figures, conditional logic, gesture recognition, and repeat loops; (ii) on a physical Pioneer 3-DX robot, confirming practical feasibility in classical navigation tasks; and (iii) in a multimodal scenario using a neural network–based LIBRAS gesture recognition module, applied to a tic-tac-toe activity for laterality training. Finally, a case study with ninth-grade students highlighted engagement, collaboration, and learning of mathematical and logical concepts. The results indicate that PINEL lowers entry barriers for beginners while preserving technical rigor for advanced applications, integrating simulation, real-robot control, and visual programming within a single environment. As such, it contributes to disseminating computational thinking and supporting reproducible, scalable pedagogical practices. Keywords: educational robotics; computational thinking ; block-based programming ; gesture recognition (LIBRAS) ; ROS 2Item Q-learning-based unmanned ground vehicle navigation in warehouse-like environments(Universidade Federal de Viçosa, 2025-03-28) Batista, Hiago de Oliveira Braga; Brandão, Alexandre Santos; http://lattes.cnpq.br/0988173500996544This dissertation investigates robot navigation in logistics environments, focusing on libraries and warehouses, using the Q-learning method. To this end, three studies are presented, each applying reinforcement learning to optimize task performance and navigation efficiency. The first study employs Q-learning to enhance book organization in the library of the Federal University of Viçosa, reducing planning time and movements by 20% compared to a greedy method while achieving a 100% success rate in task completion. Meanwhile, the second study proposes an offline Q- learning approach for unmanned ground vehicles in warehouses, outperforming traditional algorithms such as Dijkstra, A-star, and Breadth-First Search, with planning speeds up to seven times faster and a reduction in turns of up to 41%. Finally, the third study extends Q-learning to multi-agent navigation in libraries, integrating transfer learning and curriculum learning. As a result, simulations indicated a 94% success rate with nine agents, along with a 73.36% reduction in task steps compared to scenarios with only one agent. Thus, this dissertation highlights the significant potential of reinforcement learning, particularly Q-learning, to enhance robotic navigation efficiency, reduce operational complexity, and optimize logistics processes in dynamic and complex environments. Keywords: path Planning; reinforcement Learning; unmanned Ground VehiclesItem Extração automática de informações em imagens de notas fiscais(Universidade Federal de Viçosa, 2025-03-24) Souza, Mateus Fonseca de; Brandão, Alexandre Santos; http://lattes.cnpq.br/0363802323004958Este trabalho propõe uma metodologia para a extração automatizada de informações a partir de imagens de notas fiscais de energia elétrica, um processo essencial no setor elétrico, especialmente para o gerenciamento de crédito. A metodologia desenvolvida abrange diversas etapas de processamento de imagens e visão computacional. Inicialmente, é realizada a segmentação de instâncias, detecção do documento e correção de perspectiva por meio da transformação de perspectiva. Além disso, são propostos dois algoritmos para a correção de distorções de orientação; o primeiro utiliza a dilatação dos pixels de uma imagem binarizada, enquanto o segundo se baseia na Transformada de Fourier. Para melhorar a qualidade das imagens processadas, também é realizada a remoção de ruídos de alta frequência com um filtro gaussiano e realizado o realce de contornos com máscara de nitidez. A fim de garantir a qualidade das imagens utilizadas, é introduzido um método de análise de qualidade. Para tornar o processamento mais adaptável, são empregados classificadores de documentos baseados em imagens e textos. Além disso, é proposta a utilização do modelo de consultas do serviço Amazon Textract, treinado para compreender tanto o texto quanto o layout dos documentos, permitindo a extração automática de informações específicas das notas fiscais de energia elétrica. Em faturas da distribuidora CEMIG, o modelo treinado atingiu uma precisão de 0,973 no conjunto de testes. O sistema desenvolvido demonstrou ser capaz de extrair todas as informações necessárias com um tempo médio de processamento de 35,273 segundos. Dessa forma, o projeto apresenta uma solução eficaz para a extração automatizada de informações de notas fiscais de energia elétrica, com alto potencial para aplicações empresariais. Palavras-chave: Documentos; Faturas; Extração de Informações; Visão Computacional; Energia Elétrica.Item Intelligent UAV systems: payload monitoring for spraying and collaborative cargo transport(Universidade Federal de Viçosa, 2025-03-26) Barcelos, Celso Oliveira; Brandão, Alexandre Santos; http://lattes.cnpq.br/2295661594335710The use of unmanned aerial vehicles (UAVs) has been expanding rapidly, bringing innovations to various areas, such as precision agriculture and logistics. This dissertation investigates the integration of payload sensors into UAVs to improve agricultural spraying tasks and collaborative cargo transportation. Initially, a real-time payload monitoring system is proposed, which allows the UAV to adjust the application of agricultural inputs according to the variation in the weight of the spray tank. This mechanism avoids waste and improves the efficiency of the operation. In addition, an intelligent navigation system was developed that monitors the tank level and commands automatic return to base when necessary. To validate the proposal, experiments were conducted in controlled environments, demonstrating the system's effectiveness in controlling the distribution of inputs. In addition to agricultural applications, this research addresses the cooperative transportation of cargo by UAVs and unmanned ground vehicles (UGVs). An aerial transportation system was developed using a UAV equipped with an electromagnet suspended from a cable, making it possible to lift and move loads to overcome obstacles in the path of a UGV. Experimental tests have shown that the proposed approach is viable for applications such as urban deliveries, operations in remote areas and logistical support in disaster zones. The results obtained show that the integration of cargo sensors with intelligent control strategies promotes a significant improvement in UAV performance in various operational scenarios. This improvement is mainly due to the systems' ability to provide real-time information on the weight carried, allowing the UAV to make more precise autonomous decisions. This makes it possible to identify the ideal time to return to base for refuelling or to end the spraying mission, reducing waste and optimizing the use of inputs. In addition, there is more efficient management of energy consumption and flight time, based on the actual load carried. Another important advance is the automation of the load coupling and uncoupling processes, increasing agility in logistics missions and reducing the need for human intervention. These functionalities together result in safer, efficient and sustainable operations. Thus, the contributions of this work represent a relevant advance in the state of the art of UAV application in precision agriculture and autonomous cargo transportation. Keywords: Precision Agriculture; Path following; Load Transportation; Multi-robot Systems; Cable-suspended cargo; UAV-UGV Cooperation