Date of Award
Master of Science
Department of Electrical and Computer Engineering
Jeffrey D. Clark, PhD.
The U.S. military has an increased need to rapidly identify nonconventional adversaries. Dismount detection systems are being developed to provide more information on and identify any potential threats. Current work in this area utilizes multispectral imagery to exploit the spectral properties of exposed skin and clothing. These methods are useful in the location and tracking of dismounts, but they do not directly discern a dismount's level of threat. Analyzing the actions that precede hostile events yields information about how the event occurred and uncovers warning signs that are useful in the prediction and prevention of future events. A dismount's posturing, or pose, indicates what he or she is about to do. Pose recognition and identification is a topic of study that can be utilized to discern this threat information. Pose recognition is the process of observing a scene through an imaging device, determining that a dismount is present, identifying the three dimensional (3D) position of the dismount's joints, and evaluating what the current configuration of the joints means. This thesis explores the use of automatic pose recognition to identify threatening poses and postures by means of an artificial neural network. Data are collected utilizing the depth camera and joint estimation software of the Kinect for Xbox 360. A threat determination is made based on the pose identified by the network. Accuracy is measured both by the correct identification of the pose presented to the network, and proper threat discernment. The end network achieved approximately 81% accuracy for threat determination and 55% accuracy for pose identification with test sets of 26 unique poses. Overall, the high level of threat determination accuracy indicates that automatic pose recognition is a promising means of discerning whether a dismount is threatening or not.
DTIC Accession Number
Freeman, Andrew M., "Dismount Threat Recognition through Automatic Pose Identification" (2012). Theses and Dissertations. 1106.