Single frame IR point target detection based on a Gaussian mixture model classification

Abstract

This paper deals with point target detection in non-stationary background such as cloud scenes in aerial or satellite imaging. We propose a new method to estimate second order background statistics within a spatial detection method by Matched Filter (MF). Our approach consists in classifying the pixels of the image by gathering the pixels of similar covariance matrices and in estimating one covariance matrix for each class. We use a Classification Expectation-Maximization (CEM) algorithm based on a zero mean Gaussian mixture model. A key issue is the robustness of the classification with regard to target presence. This problem is efficiently tackled in a very simple manner. The resulting Gaussian Mixture Matched Filter (GMMF) has the advantage to adapt itself to second order non-stationary backgrounds and requires only the tuning of the window size used to build the observation vector associated with each pixel. Efficiency of the proposed GMMF approach is demonstrated on a large variety of cloudy sky backgrounds.