Development of high performance landmine classifier for ground penetrating radar

本研究的目的是开发一种用于探地雷达传感器的高性能地雷分类器。利用探地雷达传感器系统探测地雷的过程分为两个阶段。第一阶段是寻找阶段，所有类型的埋藏物都在这里。第二个阶段，即识别阶段，将地雷与石头、金属碎片等其他物体区分开来。在本研究中，我们提出了一种高性能的分类算法，并开发了一个用于识别阶段的简单的分类器原型。本文的主要研究成果如下：首先，提出了一种基于匹配追踪的数据预处理方法，能够准确地去除地面杂波。在去除杂波影响的基础上，从时频域提取了几种有效的地雷分类特征。接下来，我们采用了似然比检验和支持向量机这一新型分类算法作为分类方法。为了评估地雷识别的性能，我们使用数值模拟生成的数据集进行了蒙特卡罗模拟。通过仿真，验证了即使埋在浅层的地雷也能获得足够的分辨性能，并考察了土壤条件(即地表粗糙度、不均匀程度和土壤湿度)和地雷倾角对分辨性能的影响。此外，我们还从本文开发的探地雷达系统获得的测量数据中提取了特征，并验证了所提出的分类器对实际问题的适用性。

The purpose of this research is to develop a high performance landmine classifier for ground penetrating radar (GPR) sensor. A process of landmine detection using GPR sensor systems is divided into two stages. The first stage is the find stage, where all the various types of buried objects are located. The second stage, the identification stage, differentiates landmines from other objects such as stones, metal fragments, and so on. In this research, we have proposed a high performance classification algorithm and have developed a simple prototype of classifier used in identification stage. Results of the research are summarized as follows:First, we introduced a method for data preprocessing based on a Matching Pursuits and accurately removed ground clutter. After the removal of the clutter contribution, we extracted several kinds of effective features used for landmine classification from both time and frequency domains. Next, we employed a likelihood ratio test and a support vector machine (SVM) that is a novel type of classification algorithm as classification methods. In order to evaluate the performance of landmine discrimination we carried out Monte Carlo simulations using dataset generated by numerical simulations. From the simulation, we confirmed that sufficient discrimination performance was obtained even for landmines buried at shallow depths, and also checked the effects of soil condition (i.e. surface roughness, inhomogeneity, and moisture in soil) and inclination of the landmine on discrimination performance. Furthermore, we extracted the features from measured data obtained by the GPR system developed in this research and confirmed applicability of the proposed classifier to actual problems.