Target Segmentation in Scenes with Diverse Background

Abstract

Abstract. We propose a target segmentation approach based on sensor data fusion that can deal with the problem of a diverse background. Features from sensor images, including data from a laser scanner and passive sensors (cameras), are analyzed using Gaussian mixture estimation. The approach tackles some of the difficulties with Gaussian mixtures, e.g., selecting the number of initial components and a good description of data in terms of the number of Gaussian components, and determining the relevant features for the current data set. The feature selection quality is analyzed on-line. We propose a criterion that determines the quality of the resulting clusters in terms of their respective spatial distribution. The output from the analysis is used for object-background segmentation. Segmentation examples of surface-laid mines in outdoor scenes are shown.