Vehicle Detection In Worldview-2 Satellite Imagery Based On Gaussian Modeling And Contextual Learning

In this paper, we aim to study the detection of vehicles from WorldView-2 satellite imagery. For this purpose, accurate modeling of vehicle features and signatures and efficient learning of vehicle hypotheses are critical. We present a joint Gaussian and maximum likelihood based modeling and machine learning approach using SVM and neural network algorithms to describe the local appearance densities and classify vehicles from non-vehicle buildings, objects, and backgrounds. Vehicle hypotheses are fitted by elliptical Gaussians and the bottom-up features are grouped by Gabor orientation filtering based on multi-scale analysis and distance transform. Global contextual information such as road networks and vehicle distributions can be used to enhance the recognition. In consideration of the problem complexity the practical vehicle detection task faces due to dense and overlapping vehicle distributions, partial occlusion and clutters by building, shadows, and trees, we employ a spectral clustering strategy jointly combined with bootstrapped learning to estimate the parameters of centroid, orientation, and extents for local densities. We demonstrate a high detection rate 94.8%, with a missing rate 5.2% and a false alarm rate 5.3% on the WorldView-2 satellite imagery. Experimental results show that our method is quite effective to model and detect vehicles.