Author Topic: Infrared thermal video AI recognition  (Read 712 times)

0 Members and 1 Guest are viewing this topic.

Offline Dianyang

  • Contributor
  • Posts: 17
  • Country: cn
Infrared thermal video AI recognition
« on: August 23, 2018, 02:04:57 am »
Is there anyone saw these kind of video before? Will this AI algrithom suitable for infrared thermal suiveillance cameras?

Offline eKretz

  • Frequent Contributor
  • **
  • Posts: 728
Re: Infrared thermal video AI recognition
« Reply #1 on: August 23, 2018, 02:11:17 am »
I'm sure it would be fine just for notification of detection, but thermal signature would be a terrible way to ID humans. Body heat is all over the place every other minute in some cases.

Offline railrun

  • Regular Contributor
  • *
  • Posts: 64
Re: Infrared thermal video AI recognition
« Reply #2 on: August 23, 2018, 02:14:22 am »
Flir PathfindIR 2 has a pedestrian and animal detection-algorithm.

Offline Dianyang

  • Contributor
  • Posts: 17
  • Country: cn
Re: Infrared thermal video AI recognition
« Reply #3 on: August 23, 2018, 02:20:42 am »
It is an Ethernet/USB/HDMI based AI algrithom box, we are considering to combine this with some visible intelligent camera, to identify the objects by both cameras, hope this will raise the recognition rate.

Online Fraser

  • Super Contributor
  • ***
  • Posts: 8094
  • Country: gb
Re: Infrared thermal video AI recognition
« Reply #4 on: August 23, 2018, 03:13:46 am »
Not exactly what you have asked but I know of a project that used a 16 x 16 pixel people counting thermal imaging camera to identify whether the person under observation was male or female. For those unaware, male and female walking gates are different. Such gender discrimination with just a 16 x 16 pixel FPA impressed me.

Thermal imaging is a low resolution technology in many cases. Combine low resolution with wide field of view and you have low 'on target' detail. Good on target detail is needed for AI to make an identification. In facial recognition systems the distances between facial feartures is accurately measured as a sort of vector plot. Thermal imaging would struggle to present the AI with the repeatable detail required as a persons thermal profile changes according to activity and ambient temperature. Blood creates much of the thermal 'fingerprint' and the human body has a habit of redistributing it in terms of where is heated most on any given day or weather condition. The thermal profile of a face would be mainly made up of eyes and mouth being hot and nose, ears being cooler. Not much fine detail would be present for the AI to process unless the target was very close to the camera and filling most of the FOV.

Thermal imaging technology is effective at detecting a warm blooded target and potentially profiling whether it is an adult, male/female, child or animal. It detects thermal emitters that present low contrast to a conventional visible light surveillance system that is suffering low scene illumination of a camouflaged target trying to avoid detection. Thermal imaging is more of a detection rather than an identification technology, except when dealing with larger targets like aircraft where a known thermal signature helps to identify the exact airframe type. Similarly for ground or sea based vehicles/vessels.

At the end of the day, AI is only as good as the data fed to it. Poor data can lead to errors or an inability for AI to produce an output at all. Thermal imaging can be higher resolution and narrower FOV to produce more detailed identification data, but that comes at significant cost.


« Last Edit: August 23, 2018, 03:34:59 am by Fraser »

Online Vipitis

  • Frequent Contributor
  • **
  • Posts: 494
  • Country: de
  • aspiring thermal photography enthusiast
Re: Infrared thermal video AI recognition
« Reply #5 on: August 23, 2018, 08:35:56 am »
for perimiter control and detection, thermal imaging is great. hardware and software is already widely used.

for indentification, I don't it is useful. maybe a combination of multispectral images in UV, vis and thermal IR.

I believe UV would be even better for ID, as it shows so many details in the skin itself, but it will be hard to come by any large amount of training and testing data in high resolution. Phones theses days use near IR with an pattern emitter to reconstruct a depth matte.

but collect some images and feed them into the AI, you will need a lot for anything image based and that will be difficult. maybe raw values can see past weather, blood, and daytime changes and reveal underlying patterns... who knows.

Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo