Hidden target detection and classification using multiple modalities
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Hidden target detection and classification is an important task for many security and military applications. Long wave infrared (8-14 μm) cameras, otherwise known as thermal cameras, can be used towards hidden target detection and classification but are less studied in the Computer Vision literature due to their high cost and low resolution. Thermal imagery is able to reveal targets such as camouflaged or shallowly buried targets that would be hidden to optical band sensors. For this dissertation, I studied some of the problems in designing a computer vision system that uses the thermal modality along with other modalities to detect and classify hidden targets. Specifically, this dissertation seeks to address (1) calibration of multiple cameras both within the thermal modality and across modalities, (2) detection of hidden targets in the scene by identifying anomalous regions and known targets, and (3) classification of the hidden targets. I propose novel approaches towards solutions of these issues and argue for the efficacy of these approaches. Particularly, for calibration I used a ceramic backing and preprocessing technique for enhancing the contrast and its duration, and show that heating a printed calibration board is indeed viable for calibration in contrast to previous work. For detection, a dynamically updating Gaussian mixture model and sensor fusion was used to identify anomalous regions, while neural networks were used for fusing multimodal sensors and detecting known objects. Finally, for classification I developed novel thermal-based features such as water permeation and heating/cooling patterns to classify the materials. I developed the CHAracteristic Model of Permeation (CHAMP) for modeling both the rate and shape of water permeation, and use the heat equation for extracting physical material parameters for a heat feature. In each case, my results show that thermal is a useful modality for detection and classification of objects, and can be combined with other modalities to increase performance.