Implementace úloh strojového učení na mikropočítači Raspberry Pi

Abstract

This diploma thesis focuses on the practical implementation of machine learning and computer vision tasks on the Raspberry Pi 5 microcomputer platform. The introductory part presents the topic of artificial intelligence and neural networks, followed by a review of selected Python libraries and tools (e.g. TensorFlow, Keras, Pandas, NumPy) commonly used for the development and deployment of machine learning models. The work also includes an overview of available camera modules and hardware accelerators compatible with the Raspberry Pi. The practical part addresses the deployment of three solutions with different approaches to hardware acceleration — hand gesture recognition using the MediaPipe library, object detection with the SSD MobileNet model utilising the Google Coral USB TPU accelerator, and human pose estimation using the YOLOv8s-pose model deployed on the Hailo-8L neural coprocessor on the Raspberry Pi AI HAT+ extension. On top of the pose estimation, an application with two user modes is implemented — smart room management with sitting and standing classification, and ergonomic monitoring with slouching detection. In the conclusion, the performance of the individual solutions is compared in terms of frame rate and inference time, and their suitability for real-world deployment is evaluated. The contribution of the thesis lies in connecting theoretical knowledge with the practical deployment of machine learning models in an edge computing environment.