KNN classification of metallic targets using the magnetic polarizability tensor

Abstract

Walk-through metal detectors are used at check points for preventing personnel and passengers from carrying threatening metallic objects, such as knives and guns, into a secure area. These systems are capable of detecting small metallic items, such as handcuff keys and blades, but are unable to distinguish accurately between threatening objects and innocuous items. This paper studies the extent to which a K-nearest-neighbour classifier can distinguish various kinds of metallic objects, such as knives, shoe shanks, belts and containers. The classifier uses features extracted from the magnetic polarizability tensor, which represents the electromagnetic properties of the object. The tests include distinguishing threatening objects from innocuous ones, classifying a set of objects into 13 classes, and distinguishing between several similar objects within an object class. A walk-through metal detection system is used as source for the test data, which consist of 835 scans and 67 objects. The results presented show a typical success rate of over 95% for recognizing threats, and over 85% for correct classification. In addition, we have shown that the system is capable of distinguishing between similar objects reliably. Overall, the method shows promise for the field of security screening and suggests the need for further research.