I have a pedestrian dataset and would like to estimate human height in a video survillance using person detection techniques like YOLO Darknet or SSD (Single Shot Detectors). Would this technique work? Also, the videos that I have are in a constrained environment with good illumination. The idea is to get the coordinates from the bounding box and try to estimate pixel height. After getting the pixel height, some correlation could be estimated between pixel height and real world height. Note that I won't be using camera calibration.

  • $\begingroup$ You should revise your question a little bit.or else you transfer it to cross validated community. $\endgroup$
    – quintumnia
    Jan 16 '18 at 20:13
  • 1
    $\begingroup$ I did revise. Also, what is cross validated community? $\endgroup$ Mar 5 '18 at 8:02
  • $\begingroup$ Human,you should paragraph your question body! Query through the stack exchange sites,you will find it there. $\endgroup$
    – quintumnia
    Mar 5 '18 at 18:29
  • 1
    $\begingroup$ Yes, fellow human. I'll move it. Although I'm not a pro here like you, I think you should avoid using sarcasm when someone is trying to learn. No offence. $\endgroup$ Mar 6 '18 at 16:44
  • $\begingroup$ Don't take it personal,we are a community of geniuses who are eager enough to solve world problems.So let's be nice and try to keep our questions clear so as someone out there can benefit enough.Imagine if aliens cyber attack our community knowledge base! They can benefit as well. $\endgroup$
    – quintumnia
    Mar 7 '18 at 10:07

Eh, "I won't be using camera calibration." .. Not sure, what you mean.


At first, imagine a sheet of paper laying there on the floor. And first to try to transform the paper sheet (i.e. with some text on it), the picture of the sheet angled by paralaxe to the straight view: Just a transformation by a matrix. It would be highly valuable to have there some floor of square tiles. And once you would estimate/set size of the tiles on the floor, you can then transform the paper sheet.

Then, also some vertical calibration will be needed: Wherever within the scene. A wall of tiles of known shape & size would be perfect. But not really necessary: Any known/premeasured chair would be enough! ..rather two such chairs, in an opposite nooks, for numerical stability.

So yes, that could be treated as "camera calibration".


Then you can advance to the next level of the image processing:

  • Once you see feet of the person, on which tile of the floor the person stays,
  • and with the known vertical direction,

by the hid/covered known background, you can decide their height, pretty precisely, yes.

The "known background"

Just two images should be enough, of the background place: empty non-occupied, and hid/covered. The clear-empty room can be taken either onence, or periodically, or even just a partial image should be enough: just to detect shape of the person-figure.


And the third level: Really high. The person could stay in some deformed position. By interpolation of the known human skeleton to the shape of the person-figure on the screen, you can assume the curves of the skeleton of the person, so your estimated/measured total height of the person would be immune to the non-strightened angles in joints of their limbs/backbone. So even if the person knows about your observing, the AI would not get fooled. But you would need/find more refencial points on their body, to interconnect the 2D shape to the internal 3D model.

quantity of snaps in general

More snaps of the person, as they walk through the room, would be helpful: Any model, set up by any single image, should be verifiable/applicable to each of the rest of the images of the same person. Again: stability and "back testing", verifications, robustness.

You are not limited to just a single image, you can have as many as your camera is able to make. So use them, to get not only the measured size, but also to level up the "probability" of the measured result, the accuracy.

On the other hand, you can also flip the coin, and to try to guess the maximum info from the minimum snaps of the scene. But it does not seem, this would be your case. Still: Processing 24 fps recording 20 seconds long.. That definitely requires some power/time to process.


The 3D models are necessary: Of the surrounding space, of the outer shapes of the person-figure, of the internal structure of the measured body.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.