Modular interpretation of low altitude aerial images of non-urban environment

In this research project a set of computer vision algorithms for interpretation of the non-urban environment from low altitude aerial images is presented. Considering the size and spread of natural resources in non-urban areas, automating the task of gathering information about various land-covers is of particular importance. The utilization of videos captured by small aerial vehicles has many advantages over traditional high altitude aerial photography or satellite imaging for small scale environmental monitoring and agricultural applications. In this thesis the proposed Modular Interpretation Algorithm (MIA) shifts between the Coarse Tuning Algorithm (CTA), which is computationally efficient and the Fine Tuning Algorithm (FTA), which is capable of finding the target land-cover in complex situations. The CTA uses a simple colour feature in addition to contextual features of the aerial scenes for a coarse but fast extraction of land-covers in simple to medium difficulty scenarios. The FTA first uses Perceptual Grouping of Statistical Estimated Edges (PGSEE) algorithm for fitting boundaries to edge segments in aerial images. The PGSEE algorithm uses robust estimation for detection of boundary segments of elongated land-covers. Then the perceptual grouping of the estimated boundary segments used for extraction of land-covers. The 2-Stage Region Growing Segmentation (2SRGS) algorithm is proposed which uses the extracted boundaries to find the optimum position for the seed of a texture-based region growing algorithm. Finally the segmented regions are classified using LogitBoost classifier. The application of MIA for land-cover following in complex aerial sequences is investigated. The MIA inherits high accuracy from FTA and benefits from lower computational expense of CTA. The CTA, despite its lower accuracy, is still capable of finding the target land-cover in many frames and only more complex frames are processed by FTA. The computation time of MIA, despite sharing the same accuracy, is significantly lower than FTA. The combination of proposed algorithms provides a practical solution for the low altitude non-urban aerial image interpretation. The less computationally expensive CTA is used to generate a coarse interpretation of the land-covers in the image. When required, the more complex FTA is used in order to increase the accuracy of land-cover iii detection and finding land-covers that are not detectable by the simpler algorithm. The results of different stages are compared with each other as well as with a number of other competing techniques. iv To My Mother v ACKNOWLEDGMENTS