A Comparative Analysis Of The Svm And K-Nn To Detect Buried Explosive Objects Using Edge Histogram Features From Gpr Data

The Edge Histogram Detector (EHD) is an algorithm for buried threat detection (BTD) using sensor data generated by ground penetrating radar (GPR). It has been tested extensively and has demonstrated excellent performance on large real-world data sets. It has been implemented in real-time versions in vehicle mounted GPR. The EHD captures the spatial distribution of the edges within a 3D GPR volume. To keep the computation simple, 2D edge operators are used, and two types of edge histograms are computed. The first one, called EHDDT, is obtained by fixing the cross-track dimension and extracting edges in the (time, down-track) B-scans. The second edge histogram, called EHDCT, is obtained by fixing the down-track dimension and extracting edges in the (time, cross-track) B-Scans. For confidence assignment, it uses either a possibilistic K-Nearest Neighbors (p-KNN) or a Support Vector Machine (SVM) classifier. In this paper, we first propose an improvement to the EHD by adding a new feature component extracted from (down-track, cross-track) C-Scans. We show that this feature can improve the probability of detection while reducing the false alarm rate. Second, we design a large-scale experiment to compare the performance of the p-KNN and SVM classifiers and investigate their risk of over-fitting the training data. We use large datasets accumulated across multiple dates and multiple test sites by a vehicle mounted mine detector (VMMD) using a GPR sensor. The data includes a diverse set of buried explosive objects consisting of varying shapes, metal content, and underground burial depths. Performance of the EHD with the different features and classification methods are analyzed using receiver operating characteristics (ROC). To study the potential over-fitting problem, we use two different cross validation methods.