| Summary: | In the last decades, Portugal has been severely affected by forest fires which have caused massive damage both environmentally and socially. Having a well-structured and precise mapping of rural roads is critical to help firefighters to mitigate these events. The traditional process of extracting rural roads centerlines from aerial images is extremely time-consuming and tedious, because the mapping operator has to manually label the road area and extract the road centerline. A frequent challenge in the process of extracting rural roads centerlines is the high amount of environmental complexity and road occlusions caused by vehicles, shadows, wild vegetation, and trees, bringing heterogeneous segments that can be further improved. This dissertation proposes an approach to automatically detect rural road segments as well as extracting the road centerlines from aerial images. The proposed method focuses on two main steps: on the first step, an architecture based on a deep learning model (DeepLabV3+) is used, to extract the road features maps and detect the rural roads. On the second step, the first stage of the process is an optimization for improving road connections, as well as cleaning white small objects from the predicted image by the neural network. Finally, a morphological approach is proposed to extract the rural road centerlines from the previously detected roads by using thinning algorithms like the Zhang-Suen and Guo-Hall methods. With the automation of these two stages, it is now possible to detect and extract road centerlines from complex rural environments automatically and faster than the traditional ways, and possibly integrating that data in a Geographical Information System (GIS), allowing the creation of real-time mapping applications.
|
|---|