Search

In vision-based UAV navigation system, it is an important issue to identify the sub-region UAV image in the satellite image of large urban region. In this paper, we propose a method for identifying UAV image using the global feature and local feature. The global feature is obtained by the statistical analysis of lines extracted in remote sensing image. The line parameters (length, orientation, midpoint position) are directly calculated by a new algorithm based on PCA for pixels within the line support region. We also suggest the global feature descriptor based on 3D histogram of the line length and orientation. The reference images of the same size and resolution as UAV image are extracted from large urban satellite image and reference dataset is constructed from feature vectors of them. If several sub-region images are selected by searching for the global feature, the sub-region image would be finally identified by local feature among the selected images. Experiment results show that the proposed method archives a good identification rate and satisfies the robustness for the rotation and illuminate variation.

Maximum likelihood classifier is widely used in remote sensing. Many researches have indicated that using prior probabilities call make Maximum likelihood classifier perform better, but it is not true that all individual classes will have higher classification accuracy after prior probabilities are incorporated in MLC. In the present paper, the theoretical analysis about the effect of prior probability on maximum likelihood classifier is given first, including the effect on the discrimination process and classification result. Then a case study is carried out by one approach in which prior probabilities are iteratively modified by using the relative area proportion derived from the last classification as prior probabilities in next classification.