The Study Of Algorithm About Vehicle Detection Based On High Resolution Satellite Image

QuickBird satellite is a commercial high resolution optical satellite owned by Digital Globe Corporation of America, whose image resolution can reach 61 cm. Such high resolution satellite data have been widely used in land, planning, mapping, remote sensing and other fields, and especially in the field of transportation. Vehicles can be observed clearly on QuickBird satellite images, therefore people concern more and more on using high resolution satellite images to assist traffic monitoring.This paper uses a texture analysis and neural network method for the classification of vehicle targets on quickbird images. The first thing to do is to preprocess the quickbird image. The Preprocessing stage includes appropriate zooming and cropping. Through experiments and analysis, it is needed to do image enhancement on the preprocessed road clips, then this paper selects typical image region and proceeds to do texture analysis. In the progress of texture analysis, three pixel texture parameters that are sensitive to quickbird images are calculated by gray level difference statistics. The quickbird images are classified by using the feature vector that is composed of the Gray Level Co-occurrence matrix features and gray of pixel.Neural network has the capacities of self-organizing, self-learning, self-adapting and association, and it can identify characteristics of various types of samples by training the samples repeatedly. Considering the advantage of the neural network, the paper decides to extract the feature vectors as the inputs of the neural network for training.The paper finally chooses two common neural network models: Radial Basis Function Neural Network (RBFNN) and Probabilistic Neural Network (PNN) to train road and vehicles on the image. The artificial neural network is tested by using the unidentified images after the network's convergence. By analyzing the results of the 2 types of network above, it can be concluded that either network model can achieve an satisfied result. And Probabilistic Neural Network is more effective and more accurate than Radial Basis Function Neural Network as the classifier for quickbird images.