Description:
DESCRIPTION: Currently fielded explosive detection equipment uses electromagnetic signals, such as
X-rays
or MMWs to interrogate passengers and their belongings. Automatic algorithms process the images
generated by the screening hardware either to clear the passenger/property or to identify specific
anomalies for further investigation. The use of machine learning and deep learning approaches to
develop these algorithms have shown significant promise in improving overall system performance.
The DHS S&T/TSA Passenger Screening Algorithm results showed the effectiveness of deep learning
applied to passenger screening. Development of the equipment and its associated detection
algorithms is time consuming and expensive because system screening performance is difficult to
accurately model. Currently:
• Prototype systems must be built and tested to measure and understand the interaction of
X-rays/MMWs with explosives in various containment configurations.
• Development requires physical test articles to be fabricated or acquired. Suitable test
articles may even be impossible to create if the explosives involved are unsafe to synthesize.
• If machine learning or deep learning algorithms are developed for detection, many test
articles must be created and scanned to build datasets for algorithm development, training, and
testing. This is particularly labor intensive in order to generate large, representative datasets.
In order to accelerate the advancement of explosive detection equipment, the DHS S&T Directorate
seeks to develop tools to create virtual models of human travelers, their baggage and its contents.
These models:
• Should be representative of the stream of commerce.
• Should be capable of including simulated explosives and prohibited items.
• Should be able to be generated in large numbers (many thousands or millions) in a reasonable
amount of time (under 1 second per image).
• Should be useable by researchers and vendors to predict the performance of emerging explosive
detection technologies and to train machine learning-based detection algorithms. The predictions
and training will make use of tools (see, for example,
https://www1.aps.anl.gov/science/scientific-software) that simulate the propagation of X-rays/MMWs
through simulated objects.
• Should be useable for assessing a system’s ability to detect emerging threats that are unsafe
to synthesize.
• Should be useable for a variety of electromagnetic interrogation methods including synthetic
aperture radar, computed tomography, and single and multi-view (AT2) line scanners. These
technologies use transmission, diffraction, and phase contrast to detect explosives and prohibited
items.
The tools should:
• Include methods to create shape descriptions for explosives and other objects, and methods to
insert these items into representative scans. The mathematical descriptions may be based on the
union of geometric primitives, polygon meshes, and sampled three-dimensional volumes.
• Include parametric descriptions for the features of explosives, so that users do not require
access to classified information.
• Be compatible with tools in the public domain for simulating X-ray/MMW interactions with
objects.
• Be compatible with script- or code-based algorithms targeting open-source multi-dimensional
modeling software (e.g., MakeHuman and Blender)
•Provide for a real-time means of dynamic configurability, especially as regards the physical
properties of virtual materials to be used in the modeling and the system’s input/output file pathways (e.g., use of “config files”)
