Detection of buried explosive objects has been studied extensively and several sensors have been developed.In particular, ground penetrating radar (GPR) has proved to be one of the most successful modalities andmany machine learning algorithms have been developed for buried threat detection using this sensor. Largescale experiments that involved multiple detection algorithms and very large data collections have indicatedthat the relative performance of different algorithms can vary significantly depending on the explosive objects,geographical site, soil and weather conditions, and burial depth. In fact, it is possible for an algorithm thatperforms well on training data to have low probability of target detection (PD), or high false alarm rate (FAR),on new data collected in a different environment. In this paper, we investigate the possibility of developing analgorithm that can predict the performance of a discrimination algorithm on GPR data collected in differentenvironments. This can be used to select the optimal sensor/algorithm for a given location. It can also be usedto select the optimal parameters of a given discriminator for a given site. Our approach combines predictiveanalysis with adequate feature selection methods to boost PD modeling and improve its prediction accuracy.Starting from raw GPR data, we extract and investigate a large set of potential descriptors that can quantifynoise, surface roughness, and (implicit) soil properties. Our objectives are to: (ⅰ) Identify the optimal subsetof features that can affect the target PDs of a given discriminator; and (ⅱ) Learn a regression model for PDprediction. To validate our approach, we use data collected by a GPR sensor mounted on a vehicle. We extractover 50 different features from background regions and investigate feature selection and regression algorithms tolearn a model that can predict the targets PD of a given discrimination algorithm for a given lane segment. Wevalidate our results using different cross-validation methods.
展开▼
