ISEP - DM - Engenharia de Sistemas Computacionais Críticos
URI permanente para esta coleção:
Navegar
Percorrer ISEP - DM - Engenharia de Sistemas Computacionais Críticos por assunto "Accuracy"
A mostrar 1 - 2 de 2
Resultados por página
Opções de ordenação
- Deployment of ML Mechanisms for Cybersecurity in Resource-Constrained Embedded SystemsPublication . Vicente, Pedro Miguel Casal; Santos, Pedro Miguel Salgueiro dosThe increase of low security devices in the Internet is being exploited by hackers to compro mise data or use to use them as external agents to perform further attacks. As so, it is of crucial importance that networks posses a system that correctly assess the nature of incom ing and outgoing packets to protect the local network and the overall Internet connected systems. To achieve this, Machine Learning is being broadly used due to his early success. Nevertheless, these mechanisms are better inserted at the entry point of local networks, an embedded system which has limited resources to train machine learning models and/or to perform inference tasks. Since Cybersecurity is a real-time problem, the embedded systems should perform these activities in a very restricted time interval. The time required to clas sify the packets depends on the overall system load, machine learning models complexity and desired accuracy. This thesis aims to assess the current support for ML in embedded systems, either through the interoperability of models or through their development in low level languages, and the relationship between the time required by different embedded sys tems, the different tools and models. This thesis explored one transpilation tool, m2cgen, two interoperability formats, PMML and Open Neural Network Exchange (ONNX) and one real time environment, ONNXRuntime, to deploy an already trained model in a device with limited resources. Results demonstrate that ONNXRuntime was the only machine learn ing tool with a perfect match regarding samples prediction’s classification from the original models. An analysis on the time required to execute this task revealed that ONNXRun time is faster than Scikit-Learn with the Isolation Forest (ISO), One Class Support Vector Machine (OCSVM) and Stochastic Gradient Descent One Class Support Vector Machine (SGDOCSVM) models and slower with the Local Outlier Factor (LOF) model.
- Development of an Intelligent and Efficient System for Monitoring and Optimising Sailboat PerformancePublication . TEIXEIRA, HENRIQUE MANUEL DE ALMEIDA E SILVA DOS SANTOS; Ferreira, Luis Miguel Moreira LinoThis thesis introduces an AI-powered tool that improves the analysis of sailing performance by automatically detecting thin, high-visibility stripes on sails. It uses computer vision and deep learning to extract key aerodynamic parameters, such as camber, draft, and twist. These parameters are essential for understanding sail shape and enhancing performance. The motivation lies in reducing reliance on traditional manual estimation methods while ensuring efficient onboard processing with lightweight devices like a GoPro camera connected to a tablet. The research starts with a review of current computer vision and AIbased image processing techniques. It also includes a sailing-specific look at the structural and aerodynamic features of sails. Several AI methods - Feature Extraction, Line Detection, Object Detection and Recognition, and Image Segmentation - are compared in this context. The analysis finds that semantic segmentation is the best technique for the goals of this thesis. A further comparison of semantic segmentation models - SegFormer, DeepLab, SAM, and Fast-SCNN - evaluated their accuracy, efficiency, and use for real-time deployment. This review shows that SegFormer is the most effective method for detecting lines in highresolution images of a sailboat's sail. The evaluation carried out in this thesis compares a traditional algorithm, developed in a previous thesis and reused here as a baseline, with an AI-based approach that uses the SegFormer model. This implementation relies on the SegFormer mit-b1 backbone, chosen for its balance between accuracy and efficiency. Mitb2 and mit-b3 were also tested for segmentation quality and processing time comparisons. The evaluation used a dataset of 23 videos and measured how well both methods could reliably detect lines for extracting aerodynamic parameters. The results show a clear tradeoff. The traditional method consistently produced faster processing times because it relies on lightweight operations optimised for CPU use. In contrast, the SegFormer model offered more accurate and reliable line segmentation but required more computational power. Among the tested backbones, SegFormer mit-b1 was the best choice, as mit-b2 and mit-b3 resulted in significantly longer processing times without substantial improvements in segmentation accuracy. In conclusion, the traditional algorithm is still beneficial when speed and limited resources are critical. However, the AI-based approach, especially with SegFormer mit-b1, stands out as a reliable and precise option when more computational resources are available. This work illustrates the potential to integrate AI-driven computer vision into sailing performance analysis, aiding in the accurate and automated extraction of aerodynamic parameters to enhance decision-making and performance improvement in sailing.