Improving semantic knowledge of urban environment based on data fusion and machine learning methods

点击翻译按钮获取中文摘要

The first part of the project will be about understanding data fusion for complex urban environments. Data fusion can be performed at various stages of the processing pipeline, using different data and applying different methods. Due to such diversity, it is crucial to thoroughly understand the advantages and disadvantages of the methods and data used in data fusion and to select the most appropriate for further project development. An additional element discussed in this part of the work is using ontologies (grammar-based, knowledge-based) and their potential in semantic description of the urban environment elements. One of this work's objectives is to compare various tools (commercial, open-source and programming based). Therefore, it is crucial to establish measures or indicators for work assessment. Determinants of quality used in this work refer to improving the results using data fusion in terms of quality, quantity, or semantic enrichment. The measures selection will be based on the information available in literature and/or industry requirements. The indicators may refer to accuracy performance, processing time, cost reduction, range of extracted information (as data vary with the information they deliver), and the application sector. To compare different methods investigated in this project, the idea is to use mathematically simulated features (e.g., a vertical plane representing a wall) as the reference. Next, a point cloud generated from these features will be used to test how different tools (open source and commercial) and segmentation methods (based on machine learning and deep learning solutions) perform in recreating/extracting these features. This step will allow me to understand how good the reference dataset is for comparing each method against. Along with point cloud processing, several approaches will be used to detect features in image data. This will give an insight into how well data from various sensors can represent a feature and how this information can complement. The final step will be to use selected methods (with best performance based on the established indicators) on real-world data (point cloud and images). Data will represent the same area of interest to allow the fusion of information collected through the data processing. The result should deliver semantically reach information to enrich knowledge of the city environment (or selected object classes)