Vehicle Detection and Classification from Satellite Images based on Gaussian Mixture Model

Abstract

A dynamic vehicles are identified using background elimination techniques. The background removal method uses in the concept of GMM. This approach is tracking distinct feature, such as corner, edge line. The advantage of this approach is to tolerate partial occlusion, and not sensitive to image quality relative to other tracking methods, The SIFT features used here to an vehicle detection. This is done by using the images captured from the satellite. Each image is processed separately and the number of cars has been counted. This method starts with a screening of asphalted zones to restrict the areas to detect cars and thus reduce false alarms. Then it will perform a feature extraction process given by the scalar invariant feature transform in which a set of key points is identified in the obtained image and opportunely defined. Successively, using a support vector machine classifier it discriminates between key points assigned to cars and all the others. The final step of the method is focused on grouping the key points belonging to the same car to get a “one key point–one car” relationship. Finally, the number of key points finally identified gives the amount of cars present in the scene.