Automated Detection of Underwater Military Munitions Using Fusion of 2D and 2.5D Features From Optical Imagery

Abstract

Technologies that can efficiently and objectively detect, identify, and map underwater military munitions are needed. The knowledge of benthic environments adjacent to underwater military munitions is crucial for remediation decisions. When attempting to identify munitions from optical imagery, tridimensional structure information obtained from the surveyed area can complement the texture information that is available in the images. In this work, we use a fusion of two-dimensional (2D) and two-and-a-half-dimensional (2.5D) features to classify munitions on the seabed from a sequence of images of an optical survey of the seabed. The 2D features respond to texture, whereas the 2.5D features respond to geometry. The 2.5D features used were coefficients of polynomial surface fitting, standard deviation, skewness, and kurtosis of the elevation, slope of principal plane, mean and standard deviation of the distance of 2.5D points to the principal plane, surface normal, curvatures, rugosity and symmetry measures. Adding the 2.5D features increased classification accuracy relative to using only 2D features when detecting discarded military munitions.