Placeholder Image

字幕列表 影片播放

  • we have been looking at physiological monitoring to try to infer how hard people are working on a given task.

  • What we've been trying to do is find the least intrusive means for collecting physiological data.

  • We were looking at the aviation domain, but it can obviously be extended to a lot farther, not father areas.

  • So initially we had a lab study, which looked at a very simple computer task where people were tracking balls and shooting them.

  • And we could quite easily varied demand by increasing and decreasing the number of balls you know, given pattern.

  • And then we could see if the physiological measures reflected that that pattern and correlated well weighted.

  • So, um, that was a lab study, too.

  • See how feasible disease?

  • And then we went on toe doing a helicopter simulator, studying one of the questions that we had.

  • Waas would highly trained individuals and helicopter pilots reacts in much the same way as unusual person that's, you know, increased level off demand.

  • We did a computer file on something that sounds similar.

  • Where Horia was using a brain scanner and turning lights red when people were practising things like air traffic control is this a similar kind of idea?

  • Yes.

  • DEA Behind it is basically the same.

  • The Hefner sensors.

  • What was using his Maybe more accurate in certain scenarios and reflects more off the brain activity.

  • Uh huh.

  • We're aiming to use thermal imaging as a way to make it less intrusive so people don't have to wear any sort of equipment.

  • So this small flirt thermal camera the resolution is quite small.

  • It's 6 40 by 4 18 So it's not an amazing resolution.

  • There are high resolution cameras nowadays, but still not up to the level off phone camera.

  • What it does, it's picking up thermal radiation.

  • This be within the range off 7.52 I think 13 micro meters and it's converting it toe temperature.

  • So obviously that's one step off the way the other.

  • The next step was extracting temperature from various areas of the face.

  • I'm hiding behind the camera.

  • Here's this.

  • The scale here on the right, blue is cold.

  • Black is even colder.

  • And then, as you go towards reds, yellow and white, that symbolizes high, higher temperature.

  • Yeah.

  • Ah, when we extracted the data, we used the sort off images to find facial features.

  • Those facial features allowed us to get data from various areas.

  • We covered quite a large area off the face, whereas previous study looked at smaller areas like the nose forehead.

  • And indeed, actually, most of the changes occurred there on the nose and around the nose area.

  • People have reported it before.

  • We try to do, um, more structured study where we could control the demand in a more accurate way.

  • So what?

  • We've seen that usually in most people, but not in all of them knows temperatures than to drop when they're engaged in a high demand task.

  • And actually they do tend to go back up as the testament diminishes.

  • The no steep is the most evident point, but also the side areas off the nose show a similar response, just not this large.

  • Here in gray, you can see the perceived demand.

  • So there are three levels of the task towards the meat.

  • Off each of the levels, the demand peaks, things get more difficult, and at the same time you can see here in green, for example, which is temperature in point B, which was the people, the nose.

  • You can see how that drops and then, as the thus becomes easier again, it starts going back up.

  • This year in blue is Point V, which was right below the nose.

  • So here you can see also drop.

  • Maybe it's not as high, but still you can see sort of the same pattern.

  • We got a similar pattern in most people and in a lot of the helicopter pilots.

  • Some other people do have naturally colder noses.

  • They're just called by default.

  • So then there won't be any meaningful dropping temperature.

  • That's one of the one of the other effects.

  • But, you know, we're still exploring to see for those cases there are other signs indicating to the increasingly mint, Is this something that you can do in real time?

  • Or how does it work?

  • Uh, we can actually do it in almost real time, Uh, with the algorithm I used, which I have to admit, it's not highly optimized, but you could do it.

  • Maybe one friend per second.

  • The camera itself does not output a very high frame rate.

  • It's around 7.5 frames per second.

  • It's actually limited.

  • Um, you cannot buy it it higher frame rates, um, and Yeah.

  • Then the videos are quite large because it's the data is not compressed as you would get a normal visual video so it could end up being within the range of 30 GB for for one of the videos.

  • And then, yeah, there's it's quite a lot off data toe process.

  • Initially, the video is out.

  • Put it in a digital level that camera records, which is then converted toe actual temperatures by using a calibration curve.

  • And then we used that temperature basically matrix, to convert it toe visual art to be using a Kohler map which obviously can change the format coming out of the camera.

  • Is these digital?

  • So it sze the temperatures of the new convert at a video.

  • Do you ever go back from the video?

  • The visual video back to the original to get the temperature?

  • Yes, Yes, I do.

  • That audience wants to get a new area of interest and you need to extract data from it.

  • Then you go to the original matrix of temperatures and you get the temperature here from that in between those specific coordinates we're used to seeing a pixel is being three or four numbers.

  • RGB out for whatever.

  • So instead of having not, you've got a temperature and you just have a temperature.

  • While I used to convert it in sell shoes.

  • But you can convert it in any unit based on that color map.

  • We did the tracking off features.

  • We had various approaches.

  • Street.

  • The first approached from the paper that was published, uh, relied on Cascade classy fires to detect the eye area.

  • So it was trained to to pick up I patterns and to find the eyes.

  • And then we would rely on another property of the thermal image.

  • The pupils are colder than the surrounding.

  • So if you look at this rectangle which is zoomed in here, this is an actual thermal data.

  • So where seriously the RGB picture.

  • The height of this represents temperature.

  • So you could see here how temperature drops across the pupil.

  • So this is basically the center of the people.

  • The low temperature there is most likely because there's no blood vessels passing through the pupil.

  • Then he could quite accurately find the center of the pupil and then going from there, we found the rest of the landmarks for the helicopter study using this smaller camera.

  • We used a different approach which actually didn't belong to us.

  • We re trained on Calgary that was no previously published.

  • Uh, and it's more a bus to poster changes, and it works better and faster, which he trained the algorithm to pick up the specific points on the face.

  • It does fail points in extreme angles, but then we filtered out.

  • Those intervals is, you can seem, is go to the video, the nose gets colder.

  • This pilot was going through a very, very difficult scenario where he was performing an auto rotation, which means that he had no more engine power.

  • And he was just going down and tryingto cushion the landing just by using the propellers that were not speaking at the moment.

  • Here's can seem be crushed and you can see the No.

  • I was being really, really cold.

  • He looked like he left.

  • Yeah, he left.

  • Yeah, well, luckily, it was a simulator waken laugh after can your mommy Pythagoras theorem is, um mr The lights are about five seconds behind what Max is going through this much.

we have been looking at physiological monitoring to try to infer how hard people are working on a given task.

字幕與單字

單字即點即查 點擊單字可以查詢單字解釋

B1 中級

冷鼻子和熱影像 - Computerphile (Cold Noses & Thermal Images - Computerphile)

  • 2 0
    林宜悉 發佈於 2021 年 01 月 14 日
影片單字