【作者】 梅蓉；

【导师】 姜长生；

【作者基本信息】 南京航空航天大学 ， 模式识别与智能系统， 2005， 硕士

【摘要】 空战是一个复杂、动态的环境,为了在战争中取得作战主动权,必须在战斗之前彻底做好各种准备。譬如对于战场信息的获取,对战场信息的处理,对敌方目标的识别与跟踪等。如果占有了绝对的信息优势,实现了战场单向透明,就便于更快更准地打击目标。因此,基于图像信息的获取,处理和目标识别与跟踪技术成为国际上热门的军事研究课题。本论文讨论了空战中飞机目标的识别和跟踪方法研究,主要做了以下几个方面的工作: 首先分析了本课题国内外研究现状和发展方向。在此基础上重点研究了飞机目标的分割技术,提出了不同的方法进行单区域的分割和多区域的分割,对于单区域的分割,提出了遗传算法类间类内综合方差分割法和遗传算法的二维最小交叉熵分割法和彩色特征目标相似度量分割法;对于多区域的分割,提出了 Kohonen 神经网络分割法。这些分割方法均取得了不错的效果,为目标识别作好了准备。 接着研究了单目标飞机检测技术和多目标飞机提取技术。基于单目标飞机检测技术,提出了差分检测法和投影检测法;关于多目标飞机的提取技术,提出了竞争神经网络的聚类方法,为后面的识别和跟踪作好了充分的准备。 其次对目标进行识别,将目标特征样本输入到分类器中进行训练,再运用测试样本进行测试,得到良好的识别结果。主要提出了改进的支持向量机的飞机目标识别方法,线性规划的识别方法,基于神经网络的多传感器数据融合识别方法,基于飞机特征样本向量最小夹角和最小距离分类识别方法,并对目标识别效果的评价体系进行了研究。 最后简单地研究了基于颜色特征的飞机目标跟踪方法。提出了分割彩色区域来更新彩色模板,并运用卡尔曼滤波进行跟踪预测,从而达到跟踪的目的。更多还原

【Abstract】 Airfight is a complex and dynamic environment. To get initiative droit during airfight, it is necessary to make different sufficient prepration before airfight, for example, information extraction, information propcession, enemy target recognization and tracking technique, and so on. If one has played absolute importance on information to implement monomial clarity of war field and more quickly and truly beat enemy targets. So the technique of information extraction, information propcession, target recognition and tracking has become internationally pop military study task. This paper mostly has discussed recognition technique of plane targets at airfight, it mainly includes the following several aspects: Firstly, analyse study actuality and development direction of this task. This paper mainly studies segmentation technique of plane target, providing different ways to single area segmentation and multi-area segmentation. On single area segmentation, it provides 2-D minimum cross entropy of dynamic image segmentation based on genetic algorithm and variance between two classes and variance in two classes based on genetic algorithm and the segmentation of similar measurement of color character targets;on multi-area segmentation,it provides the segmentation of Kohonen neural network. These segmentations all get good effect to make preparation for the following recognition. Secondly, this paper gives single target plane detection technique and multi-target plane detection technique. On single target plane detection technique, it provides difference detection and projection detection; on multi-target plane detection technique, it provides clustering way on competition neural network, which make full preparation for the following recognition and tracking. Thirdly, to recognize target, is to train target character samples imput to interpolator, then test using testing samples to get recognition result. It mainly provides the plane recognition algorithm of improved support vector machine, the recognition algorithm of linear programming, the recognition of multi-sensor data fusion on neural network, the recognition algorithm of the minimum angle and minimum distance between vectors of plane character sample, and studies the appraising system which infects recognition effect. Finally, simply studies plane target tracking technique on color character. It provides segmenting color area to update color template to track targets.更多还原