| Abstract [eng] |
In the last decade, machine learning has been dominated by so-called deep neural networks, which utilize rapidly developing computing and data processing technologies. The aim of this work is to identify, investigate and adapt the best architecture of the deep neural network for parking lot occupancy. Initially, a literature review was carried out to review the various regional-based architectures of deep neural networks and to briefly introduce color spaces that are often adapted to solve photo processing tasks. The methodological part provides an overview of the networks used in the research, their structure, and the possibilities of using them by applying the knowledge transfer method. In order to compare the networks of different architectures during the research, the method of comparison of average precision networks, often used in literature, was taken into account. The project part focuses on the hardware used, the process of preparing various data sets and the training sequence of the networks. Introducing the most relevant network and color space search results. Network adaptation review and speed analysis are introduced. The results obtained are comparable. Finally, based on the findings of the research, an algorithm for marking and calculating parking spaces was designed. |
