IPVM shows new IR body temperature cameras using biometric face detection systematically unreliable
An investigation by IPVM into camera-based fever screening systems has discovered widespread manipulation of results, undermining the accuracy and safety value of the devices, which have been widely implemented with biometrics to fight the spread of COVID-19.
The study has been published in the Journal of Biomedical Optics which shows that body temperature measurements with thermal imaging systems released by a number of companies in 2020, often integrated with biometric access control systems, are not only inaccurate but actually misleading.
Part of the effectiveness difference is a choice to use lower-resolution IR sensors, according to the researchers.
“Newly released IRTs commonly use IR sensors with much lower resolution than FDA-cleared IRTs,” the study states. “For example, Heimann sensors are a widely used component “temperature tablet” IRTs, which are available in resolutions ranging from 16×16 to 80×64pixels; sensors with a resolution of 32×32pixels or 1024 pixels in total, are the most common.12 By comparison, the FDA-cleared FLIR E54 (a control device in this study) uses a sensor of 320×240pixels or 75 times as many pixels at 76,800 pixels.”
Target selection also tends to work differently. While all use biometric face detection, traditional IRTs target the inner or “medical” canthus (the corner of the eye), while the newer systems tend to measure temperature at the warmest area of the face.
IPVM tested five tablet solutions; the Bems Temperature Terminal, the TVT TD-E2128-TM, the Certify SnapXT Pro, the Meridian Clear 2, and the ZKTeco 8” SF1008+, as well as bullet cameras Dahua DH-TPC-BF5421-T and the Hikvision DS-2TD2636B-13/P. Each was introduced in 2020, went to market without an FDA 510(k) regulatory review, and is either known to be used or likely to be used.
Amazon invested $10 million in Dahua thermal cameras to keep its shipping facilities operational during the pandemic, though a union representative says there have COVID-19 outbreaks at 30 Amazon facilities across North America.
Each device was tested at different temperatures mimicking low, normal and elevated body temperatures, with each measurement also noted by two control devices, one of which was a handheld NCIT.
IPVM found that all seven tested devices systematically reduce temperatures above 37 degrees Celsius and raise temperatures below 36 degrees Celsius, and tend to record temperatures in a normal human range of between 36 and 37 degrees. The FDA-approved control devices had near-perfect temperature accuracy, in contrast.
The Bems device nearly always recorded a normal temperature, regardless of the subject’s temperature setting. None of the tablet devices had a regression coefficient above 0.6, and the bullet cameras regression coefficients did not top 0.8 (a coefficient of 1 indicates a perfect correlation). Rather than random or constant systematic errors, the tested devices clearly show bias towards normal human temperatures, overestimating low temperatures and underestimating high ones.
These results occurred in uniform laboratory conditions, without the challenges of subject movement, inconsistency in distance, and other variables that make the real-world task significantly more difficult.
The handheld device, interestingly, was also found to be significantly inaccurate, but in a consistent way, with subject temperatures tending to register higher by around a half-degree Celsius, and never delivered a reading lower than the subject’s temperature.
When contacted for comment by The Washington Post, a Certify representative denied that the company’s IR body temperature scanners alter high temperatures, but said a failed reading would return a default measurement of 96 degrees. Representatives of Dahua, Meridian and ZKTeco told the Post their systems do not manipulate temperature measurements, but in some cases “self-calibrate” based on their environment with help from their software.
“The deviation setting is not intended to ‘distort’ results. It’s designed to allow customers to receive alerts only when actual threats exist,” ZKTeco CEO Larry Reed said.
The FDA has already been warned of the problem in a letter from the Security Industry Association, written in collaboration with IPVM.