İNSANSIZ HAVA ARACI GÖRÜNTÜLERİNDEN KENTSEL ALANLARDA ARAÇ TESPİTİ
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
It is very important for the planning, management and sustainability of urban areas, especially in metropolitan cities to automatically detect and analyze the change of objects such as buildings, trees, and vehicles using satellite images or aerial photographs with various methods. Obtaining this information with classical methods such as terrestrial measurements causes a lot of time, cost and labor loss. Hence, it is important that the work done in this area is detected by automatic or semi-automatic methods using satellite images or aerial photographs. In this thesis, an approach has been developed for the detection of stationary vehicles from very high spatial resolution color and three band (Red, Green, Blue) images obtained by unmanned aerial vehicles (UAV) in urban areas. Images were taken with a UAV at the Beytepe Campus of Hacettepe University in the study. In this study, first, a digital surface model (DSM) was generated by image matching and automatic correlation technique followed by orthophoto production. Then, three test fields (Test Area # 1, Test Area # 2 and Test Area # 3) with different characteristics were selected from the orthophoto of the whole study area. In the next step, multiresolution segmentation followed by supervised classification was performed using three band (RGB) orthophoto data and elevation data as an additional band. Then, a reference dataset in vector format was created by drawing a closed area over the outer boundaries of each stationary vehicle in the test fields from the orthophotos. At the last stage of the work, stationary vehicles determined by the proposed method and the reference dataset are overlaid and accuracy analyses are performed. In this context, vehicle detection percentages and quality percentages are calculated and reviewed by considering the accuracy values in three different categories as True Positive (TP), False Positive (FP) and False Negative (FN). According to the obtained results, the vehicle detection percentage for test area # 1 is 88.99%, the quality percentage is 51.56%, the vehicle detection percentage for test area # 2 is 78.53%, the quality percentage is 55.17% and the vehicle detection percentage for test area # 3 is 92.15% calculated as 72.43%. It has been observed that the heights of nonvehicle objects such as buildings and trees in test areas influence accuracy analyses. In particular, the stationary vehicles parked in close proximity to each other and the ones that are surrounded by the trees and parked under the roofs of the buildings are affecting the results negatively. It was observed that the accuracy of the DSM and the point density directly affected the vehicle detection percentage. Hence, it is expected that an increase in the spatial and spectral resolution of the orthophoto as well as an increase in the accuracy of the DSM will increase the vehicle detection percentage and the percentage of quality values. Obtained results show that automatic detection of stationary vehicles from very high spatial resolution RGB images can be performed with high accuracy using the method proposed in this thesis study.