The goal of this work is to take advantage of general environments (e.g., Orange) to efficiently build machine learning models, and adapt them for implementation in embedded systems. The reason is that the latter have limited resources and adaptation of models is often not a simple task. General environments usually do not support this adaptation of models. However, there are dedicated tools from embedded systems vendors to create customized models, but these tools are much more limited in terms of functionality and choice of models. Therefore, we tried to take advantage of both types of development environments. The models were created with the general tool Orange and then adapted for effective use in embedded systems with different approaches using the dedicated tools mentioned above. In the process, we wrote several different programs in Python and C. The contribution of our work are program codes and sequences of steps that facilitate the creation of machine learning models in the Orange environment, and their efficient adaptation for use in embedded systems. In this way, the procedures can be performed by users with less experience in machine learning. We tested these procedures on two practical examples for classification and recognition, and analyzed the effectiveness and accuracy of the models in all steps from the general environment to their application on embedded systems. Our tests have shown that the models perform comparably well when applied to more limited embedded systems. All the work done in this thesis is publicly available on the GitHub repository [11], which is available for further extensions.
|
