| Popis: |
This paper describes a suite of target cueing algorithms which has been developed for the recognition of ship targets in the open ocean through FLIR Imagery. Imaging prepro cessing is first used to remove pattern and temporal noise. A relaxation technigue implemented to extract the target's silhouette. The superstructure profile is then obtained and classification is performed based on low-order coefficients of the discrete Fourier transform of the profile. This classification approach was found to have a 93% accuracy for short ranges (7-11 miles) and 70% long (11-20 for eight target classes tested against 11398 images. Finally, a terminal homing algo rithm is described which incorporates scene tracking for maintaining track on a selected aimpoint which demonstrates superior performance over more conventional approaches.The purpose of this paper is to present the results of a Ford Aerospace effort in ship target classification and aimpoint maintenance. A distinctive attribute of the present study is the utilization of an extensive data base provided by the Naval Weapons Center, China Lake, California. This data base represents a unigue manifestation of careful data gathering and ground truthing.The classification phase of this paper presents a template matching approach based on an aspect angle versus range binning of the data and a correlation metric matching algorithm. The aimpoint maintenance part of the study presents an algorithm which ascertains the global scene motion in a frame seguence thus allowing aimpoint track in spite of scene perturbations such as translation, rotation and scale changes.The present study does not address the detection problem aimed at gathering potential targets or target-like objects through a search pattern over a wide f ield-of - view.Once a potential target region has been localized, it is the goal of the segmentation algorithm to extract the target pixels from their immediately surrounding background. In the case of FLIR imagery of ship targets on the open ocean, the image can be simply modeled as being two-class, since in most cases the target is either brighter or darker than its background due to the thermal properties of the scene. Conseguently , a two- class relaxation segmentation algorithm^1) which extracts the target based on pixel intensity has been developed for use with the ship imagery. This algorithm uses a three-step approach to segment the image. In first step, initial separation of target from the background is achieved based on gross intensity differences. In the next step, relaxation updating of the pixel labels is applied which uses the spatial neighborhood consistency to remove any ambiguity in the labeling of pixels. This relax ation step allows the algorithm to produce consistent segmentation results despite the high noise levels which often characterize FLIR imagery. In the final step the pixel labels are interpreted in order to form a binary image of the target silhouette. From this silhouette, a geometric target outline is extracted from which features for classi fication can next be defined. Because the water line of the ship is often suspect due to variations in loading and ocean conditions, most of the shape information of ship targets is derived from their superstructure. |
