Recognition of Vehicles as Changes in Satellite Imagery

Recently, a new probabilistic representation for 3-d volumetric appearance modeling (VAM) has been developed. In this thesis, VAM is used to model background appearance of a scene given by satellite imagery. The vehicle motion constitutes a significant source of change in this imagery and it is possible for human analysts to identify individual vehicles at the resolution of ∼0.6m x 0.6m per pixel. In this thesis, the aim is to automate this process by characterizing changes given by VAM as vehicles in new images of the scene. Despite the improvement in terms of satellite imaging standards, the resolution on the vehicles in this imagery is rather poor compared to the resolution of the objects typically studied in object recognition literature. A typical vehicle of size 1.7m x 4m covers 2.5 pixels by 6 pixels in modern satellite imagery. In this thesis, a vehicle recognition framework that operates at this extreme resolution is proposed. The vehicles are represented as a composition of primitive parts given by an oriented blob detector. The blobs are detected as intensity extrema given by a second order derivative of a Gaussian kernel. The normal appearance model of the scene is utilized so that primitives are detected on the foreground objects with more likelihood. In the training stage, a compositional part hierarchy is learned to represent the geometry of primitive parts exhibited by vehicles. In the test stage, the learned compositions of change primitives are detected and labeled as vehicles in the scene. The performance of the proposed method is measured on two datasets cropped from Quickbird satellite imagery covering highways, inner city roads and parking lots as well as other areas. The detection accuracy is significantly improved over the baseline performance of the change map given by VAM. Three other detection algorithms applicable at the resolution of the application are also implemented for comparative performance analysis. These are template matching, PCA and PCA-SIFT based methods and all three exhibit worse performance than the proposed method.