T&E of the SVI Face Capture & Matching System (Final Report)

• NCJ Number: 249541
• Date Published: November 2012
• Length: 71 pages
• Grant Number(s): 2010-IJ-CX-K024
• Agencies: NIJ-Sponsored
• Publication Type: Test/Measurement; Research (Applied/Empirical); Report (Technical); Report (Study/Research); Report (Grant Sponsored)
• Award: Sensor, Surveillance, and Biometric Technologies Center of Excellence

Annotation

This report describes the methodology and presents the results of the testing and evaluation (T&E) of the StereoVision Imaging (SVI) Inc. prototype, which consisted of a binocular imaging subsystem and a laptop that hosts the imaging and face-matching subsystem.

Abstract

The T&E focused on the ability of the SVI software, FRT Tracker, and Disparity Calculator to extract or segment the faces contained in captured images. The FRT Tracker software is automated to extract the images from an SVI binocular connected to a laptop and provide them to the LI ABIS subsystem for facial recognition. The images from the Disparity Calculator are provided as an input to the LI ABIS. The T&E produced six main conclusions. First, the Disparity Calculator failed to produce occluded (facial images without background surroundings) that were acceptable in quality to the LI matcher. Second, the change in match performance between SVI images collected at 50 meters using the preset 1 and preset 3 is not statistically significant. This led the test team to believe that the SVI binocular images are not correctly focused. Third, there is a stark improvement in match performance produced by matching all occluded (Run 6, 27.4 percent) compared to running all left (Run 4, 1.0 percent) and right (Run 5, 0.7 percent) images produced by the Disparity Calculator. Fourth, The improved match performance of the MegaMatchers over the LI matcher in three runs led the test team to believe that some parameters of the LI matcher are incorrectly set. Fifth, The Disparity Calculator produced occluded images that often contained partially or fully obscured faces. Sixth, the Sony DEV-3 binocular did not produce face images at 50 meters or greater for the purpose of automated face matching. 50 figures