Machine Vision Resources for Smart Fever Detection Systems

Side-note: The smart glasses pictured above are using a Flir brand thermal camera, but is sold by:
ARSENZ ThermoGlass   in Taipei City   LWIR Technology,   80×60 @9FPS,  51° FOV (25° option)
  (Could make for a stealthy Fever Inspector if the camera was just accurate enough)

The Proposed Approach

From  OpenCV: Facial-Feature-Detection Tutorial 2015 tutorial

Use machine vision to identify a face, or a particular area on a face to measure.

In the application of a single point IR sensor with long range optics:

Use a servo driven Gimbal to point the sensor at your area of interest.

RC Servo Driven Camera Gimbal

This Gimbal assembly, or just the plastic bits are widely offered by Chinese sellers, this example is currently offered for $4.61 delivered in > a month.  You can get the identical thing from an Amazon warehouse in a few days for $17.95

I don't know what the mount was originally for, perhaps a drone camera. These RC servos can be directly driven by an Arduino computers.

Below are some examples of a single point IR sensor being scanned over a scene to generate an image, I'm not proposing doing that for fever detection, rather an AI system would point the sensor to a single point for a reading.

Here's an example application generating 42x32 image by scanning a thermometer gun sensor over a scene with servos, the 3072 points image requires a 2 minute scan.

This demonstrated setup could be aimed by an added AI machine vision camera.

Here's a second servo scanned thermal imager, using the often discussed  5° FOV MLX90614ESF-DCI sensor, Schematic, Arduino (& Javascript?) source code is provided. There is also a Detailed Construction Tutorial.

His Demonstration Video

Here is yet another servo scanned thermal image generator

Using Mathematica to do it


If using a thermal camera with modest pixel resolution in an automated system it also may have to be pointed at the subject in order to get usable spacial resolution.  I though a thermal camera that could screen your visitors would be a nice addition to a Ring Doorbell camera.  But some of these camera have a 180° FOV, trying to match that coverage with a typical thermal camera would only have a few pixels for a distant face.  Aiming a telescopic thermal can on an individual face would not better resolution.

The Face Recognition.

 Some OpenCV Tutorial Examples (Pay for code DL)

Face Recognition: An Introduction for Beginners

Face Detection – OpenCV, Dlib and Deep Learning ( C++ / Python ) (4 techniques)
Facemark : Facial Landmark Detection using OpenCV

 

A brief distraction,  here's a demo of the YOLO (You Only Look Once) object recognition algorithm innovation, running on a powerful desktop computer with a gaming coprocessor (I assume) this is the sort of processing power a self driving car needs.

Yolo 2 analysis of a James Bond Scene

Small Machine Vision Boards

There are a variety of devices that can do Machine vision, it requires significant horsepower (the more the better) it can't be done by an Arduino like microprocessor.

Demonstration of the $21.90 Maix computer using the innovative Kendryte K210 module with Inference engine,  The demonstration is run from their micropython implementation.

 (note the dots identifying facial features)

 Sipeed Maix Kendryte processor boards

The $21.90 Sipeed Maix  computer and it's variations with  built in inference engine can be found at Seedstudio.com the 230 Giga operations per second inference engine is a modest performer.
                                Other fueatures: , dual Risc-V processors, hardware FFT, and built in LCD driver.


In my experience it does fine with well lit printed images but the camera may not be sensitive enough for typical real world lighting conditions.

 Kendryte K210 processor module $7.90,  $8.90 with WiFi
 (it was designed by the largest supplier of Bitcoin mining ASIC processors)

The OpenMV Cam H7 $65.00

This newest hardware incarnation had made a big jump in processing power, using one of the most powerful (480 Mhz) Arm processors available.  It uses the python counterpart of the better known OpenCV machine vision library.  The Sipeed-Kendryte boards are running an implementation of OpenMV.

There are clones of this open source Hardware offered for a variety of prices.

 

The Jovois Machine Vision Processor-Camera $49.99  

"Open-source quad-core camera effortlessly adds powerful machine vision to all your PC/Arduino/Raspberry Pi projects."

Kickstarter sales demo jevois.org      Store and newer info jevoisinc.com

 The 6 minute demo

This tiny module designed by a UCLA ? professor, it's big advantage is running Linux (like a Raspberry Pi) which enables it to run a large variety of machine vision code, it's easier to experiment with than some other devices, although the operation is a bit quirky, you need to connect it to your computer and use certain video player programs, selecting a resolution setting to change Vision Processes.

Linux based systems are probably not as well suited to embedded processing applications, (slow starting, brown out sensitive)

My Cheap Gimbal experiments, the example on the right uis using the cheapest plastic gear servos, everything is supported by the shaft of the bottom servo, the setup on the left is using metal gear servos, costing about $1 more. I had to fabricate a rather oversized mount to fit the Jovois in it's awkward orientation of the camera makes for an offset load, but more importantly the pivot point  (blue arrow) would ideally be at the focal plane of the camera.
 In this sort of application a serial interface on the sends readings to an Arduino driving te servos.

Jevois Object Tracking Demo (Not exactly what we want, maybe just the sensor on the Gimbal)

Target tracking car with swivel head vision (Not very relevant to out goals) ArUco markers

 

 

 

 

Raspberry Pi and Intel® Neural Compute Stick 2

An afterthought addition that can run OpenVino which can be scaled up to more powerful Intel hardware including CPUs, iGPUs, FPGAs, VPUs as needed.   See the example OpenVino apps

 The chipset can hit 4 teraflops of compute and 1 trillion operations per second of dedicated neural net compute at full blast.  (Compared to 230Gigops ( fixed point op) in the Kendryte chip)





Google AIY Vision Kit  $89    (Target, Adafruit, Mouser)

uses the Intel processor ver.1 with a Raspberry Pi  Zero













  A slightly smaller form factor Raspberry Pi is the Pi3A, the A models have a slightly smaller footprint th may be useful in embeddd applications, but less connectors and ram, it is also $10 cheaper.  (The raspberry Pi Compute modules are smaller yet and may be the only Pi's designed for commercial use)

Raspberry Pi 3B



 The FLIR New Products at CED 2018 Vvideo showing those smart glasses pictured at the top of this page.

Here's a paid design service advertising thermal body temperature scanning device concepts.


Palm sized Video and Infrared Camera Gimbal for Drones (when money is no an issue) 2015
                                         (Note the milky white window for the IR camera.)

Drone Harassing Chinese Not Wearing Masks (audio seems to be reworked)

A fever detecting AI police helmet, said to work from 5 meters away,  got broad coverage in the press, after some effort I believe I found the camera used, a USB cell phone accessory with 2% accuracy, a bit ambiguous with the "normal" and "take him away" temperatures are just 1.8°F apart.

Heads up display view

 They don't look identical to me