Analysis of infrared polarisation signatures for vehicle detection

Abstract

Thermal radiation emitted from objects within a scene tends to be partially polarised in a direction parallel to the surface normal, to an extent governed by properties of the surface material. This thesis investigates whether vehicle detection algorithms can be improved by the additional measurement of polarisation state as well as intensity in the long wave infrared. Knowledge about the polarimetric properties of scenes guides the development of histogram based and cluster based descriptors which are used in a traditional classification framework. The best performing histogram based method, the Polarimetric Histogram, which forms a descriptor based on the polarimetric vehicle signature is shown to outperform the standard Histogram of Oriented Gradients descriptor which uses intensity imagery alone. These descriptors then lead to a novel clustering algorithm which, at a false positive rate of 10−2 is shown to improve upon the Polarimetric Histogram descriptor, increasing the true positive rate from 0.19 to 0.63. In addition, a multi-modal detection framework which combines thermal intensity hotspot and polarimetric hotspot detections with a local motion detector is presented. Through the combination of these detectors, the false positive rate is shown to be reduced when compared to the result of individual detectors in isolation.