Segmentation and classification of man-made maritime objects in TerraSAR-X images

Spaceborne monitoring of wide maritime areas can be suit able for many applications such as tracking of ship traffic, surveillance of fishery zones, or detecting criminal activities. We present novel approaches for segmentation and classification of man-made objects in TerraSAR-X images including estimation of orientation and size. This is a difficult task as detections are affected by clutter and noise effects, and each object can have different appearances. We chose a statistical approach to robustly segment given detections using Local Binary Pattern (LBP) and Histograms of Oriented Gradients (HOG). This is the fundament for subsequent feature analysis and 3-stage-classification based on Support Vector Machines (SVM) with separation of clutter and man-made objects in first, non-ships and ships in second, and different ship structure types in third stage. An experimental evaluation demonstrates the effective operation of our approaches.