Terrain segmentation for safe landing zones for UAVs

Sammanfattning

This thesis expands on a developed method for aerial image segmentation. The method utilizes feature values derived from color and texture to classify regions into a set of predefined terrain types. The aim is to enable an unmanned aerial vehicle (UAV) to sort the surrounding surface below it into regions of interest, identifying safe landing zones or areas to avoid. Feature extraction are executed through the use of chromatic co-occurrence matrices between the RGB and HSV color channels, gray-level co-occurrence matrices (GLCM) and local binary pattern (LBP) approach. The approach is constructed to work in two phases. First is the learning phase, where the algorithm receives sampled data of the defined terrain classes. it extracts feature values and analyzes them to identify distinguishing characteristics for each class. The second is the segmentation phase, where the algorithm uses the characterizing features found in the learning phase to classify and segment an aerial image through a probabilistic voting scheme. The images used for sampling and evaluation are provided by a semantic drone data set provided by the Institute of Computer Graphics and Vision (ICG). This thesis work did not result in a fully robust terrain segmentation system. A potential weakness lies in the analysis of the extracted data, which may have led to a biased and unreliable voting scheme. Improvements could include: providing a larger and more consistent number of samples, removing eventual biases caused by that aspect. Additionally, incorporating artificial intelligence (AI) could streamline the analysis process and improve overall efficiency.