Landmine classification using possibilistic K-nearest neighbors with wideband electromagnetic induction data
Metadata[+] Show full item record
[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] In this thesis, a possibilistic K-nearest neighbor classifier is presented to distinguish between and classify mine and non-mine targets on data obtained from wideband electromagnetic induction sensors. The goal of this work is to develop methods for classifying wide-band electromagnetic induction data into one of several target classes or a non-target class. For some landmine detection systems, it could be necessary or helpful to discriminate between the several classes of targets, so that they can be analyzed and processed according to their specific properties. For example, it might be of importance to distinguish mines with high-metal content versus low-metal content or distinguish between anti-personnel versus anti-tank landmines. The proposed classifier achieves this goal using a method that is motivated by the observation that different buried object types often have consistent signatures depending on their metal content, size, shape, and depth. Given a sparse representation obtained using the joint orthogonal matching pursuits algorithm, particular target types consistently selected the same dictionary elements for their sparse representation. The proposed classifier distinguishes between particular target types using the frequency of dictionary elements selected by an alarm. Possibilistic weights are assigned for each alarm for sixteen landmine target classes as well as a false alarm class. The proposed classifier accuracy is compared to several state of the art methods and it shows improvement in discrimination results.
Access is limited to the University of Missouri - Columbia.