Advanced Automated Threat Recognition in Security Imaging [INACTIVE]
R4-B.3

Download Project Report (Phase 2, Year 3)

Project Description

[CONCLUDED JUNE 2016]

Overview and Significance

This project is organized along two thrusts. The first thrust is high-throughput screening. We propose algorithms based on a hierarchical network of classifiers. This is critical both in portal systems, where high throughput requires significant automated decision support, and in standoff systems where the proliferation of multimodal data can overwhelm human interpretation. Specifically, this project will leverage existing sensors, imaging modalities and explosive detection algorithms.

Some modalities such as Active Millimeter Wave (AMMW) and Human Inspection can be time-consuming. To improve detection performance and maintain high-throughputs, the proposed scheme will selectively route subjects sequentially through different stages. Subjects that do not pose threats exit the system early. In our preliminary experiments involving several benchmark datasets, we have shown that on average our scheme can improve throughput by as much as 50% without sacrificing detection performance. We have also conducted experiments with AMMW, Infrared and Passive Millimeter Wave (PMMW) modalities. For this scenario in our proposed scheme, we can show that with 47% AMMW utilization, namely, on 47% of selectively chosen subjects, we can match the detection rate when AMMW is used on all of the subjects. Since AMMW is far more time-consuming than Infra-Red (IR) it follows that our throughput gains can be significant.

The second thrust focuses on detection methods. Our high-throughput screening methods hinge on high-accuracy detectors for different imaging and sensing modalities. The goal in this thrust is to develop concepts for next-generation ATR algorithms for Advanced Imaging Technology (AIT) and standoff detection that are robust, increase probability of detection, reduce false positives, and extend to broad classes of AIT/ATR sensors, such as mm wave scanners, passive IR, X-ray backscatter, and other concepts. The suite of new algorithms will improve effectiveness in the screening of people in airports, which is of significant interest to TSA. Our fundamental assumption is that it is too slow or costly to collect full sensor data on every object of interest, either for training, or during real-time operation. Thus, it is important to develop technologies that identify the right set of information to collect on individuals automatically, based on the most recent collected information.

The new algorithms will also impact standoff detection algorithms for suicide bombers, which is of interest to other agencies within DoD and the State Department. The efforts are aimed at developing a robust surveillance system for pervasive and persistent detection capability. Improved ATR concepts for AIT is of particular interest to mm wave and x-ray backscatter vendors. Our goal is to perform standoff detection of concealed explosives at low false alarm probability and near certain probability of detection. The long-range impact of this research will be the development of adaptive, high throughput risk-based screening algorithms for different combinations of sensing modalities that exhibit improved sensitivity/specificity over conventional approaches.

Our work is closely related to the prediction time active feature acquisition (AFA) approach in the area of cost-sensitive learning. Our objective is to make sequential decisions of whether or not to acquire a new feature to improve prediction accuracy.
Phase 2 Year 2 Annual Report
Project Leader
  • Venkatesh Saligrama
    Associate Professor
    Boston University
    Email

Faculty and Staff Currently Involved in Project
  • Joe Wang
    Doctoral Candidate
    Boston University
    Email

Students Currently Involved in Project
  • Jing Qian
    Boston University