Your problem seems to consist of two parts. First, you want to determine which pixels belong to the person, so that you can measure their height in pixels. Then you want to determine how the pixels correspond to physical distance. Knowing no details of your images (e.g. how complex is the background), I will not attempt to answer the first part - in worst case, you can even manually draw a bounding rectangle around the person.

Computing physical distance out of pixels requires you to perform camera calibration and this is the subject you might want to look up. There are several aspects to this (way more than fits a single answer), but most basically you want to "rectify" your image to account for camera perspective. To do that, you want to find the camera matrix. The usual way of doing that (e.g. in OpenCV which I'm familiar with, but there should be corresponding functions in Matlab) is to calibrate using a chessboard pattern: if your person was holding a chessboard of a known square length, there are built-in functions in OpenCV to detect this pattern and compute the camera matrix for you. Read more here:

https://docs.opencv.org/2.4/doc/tutorials/calib3d/camera_calibration/camera_calibration.html

All is not lost if you don't have a chessboard, though. You can still calibrate if the person is holding an object of known dimensions. This lets you find 4 points on the image about which you know (1) their coordinates in image space (pixels) and (2) their coordinates in some physical space (millimetres). You can use getPerspectiveTransform to compute the transformation from pixel space to physical space, effectively obtaining a camera matrix.

https://docs.opencv.org/3.4.3/da/d54/group__imgproc__transform.html#ga8c1ae0e3589a9d77fffc962c49b22043

Once you have the camera matrix, there are two ways to proceed. You could transform your entire image using warpPerspective and then work with the resampled image as if each pixel was exactly one millimetre (or centimetre, or whatever you defined your physical coordinate system). Alternatively, you could work with the original image and extract only the most interesting points (e.g. corners of a bounding box around the person), and then just transform their coordinates by multiplying their coords with the camera matrix.

https://docs.opencv.org/3.4.3/da/d54/group__imgproc__transform.html#gaf73673a7e8e18ec6963e3774e6a94b87

Of course, there's more to read and do in terms of calibration, you might for example want to correct for the distortion of the camera lens. But I hope the above will serve as a good primer.

PS: Your EXIF data will not help you at all. Knowing the focal length does not help since you don't know the precise location/angle of the camera with respect to the observed scene.

Thanks for your answer. Unfortunately I have Matlab and cant seem to find a way to do calibration w/o a checkerboard. There is https://de.mathworks.com/help/vision/ref/estimategeometrictransform.html#btrzgde-transformType which seems the same. But using a 'projective' transform doesnt seem to work. I picked 4 points of the known object and matched them to the world dimensions of the known object (0,0; width, 0; width, height; 0, height;)

You have to know the distance from you camera and the frontal plane. This you can have an easy test by positioning a grid paper at the distance you will take a shot. When you know the grid size (1cmx 1cm grid) and pixel size - you can calculate the ratio, and verify the calculation. So, in that way you can learn how to calculate the size of an object taken at particular distance. If you assume ideal objective, you can use proportional sizes for different distances.

Viara,

if you compute scale ratio between the physical dimensions of some known grid and its pixel size, you will come up with a simple factor that is however not very accurate. The problem lies in perspective transform: objects are projected from world coordinate system onto the image plane, which means those which are closer to the plane will appear larger. Rupert wanted the most accurate method to determine physical shape, and camera calibration is the way to achieve this. Knowing the focal length and a few other parameters would allow one to compute the intrinsic parameters matrix explicitly, although I don't know if this is beneficial over estimating its parameters from calibration images. See the following on this subject:

https://www.mathworks.com/help/vision/ug/camera-calibration.html

Rupert,

I don't know Matlab's CV toolbox well enough to help you with this directly, but I'm certain this can be done. It's MATrix LABoratory after all, and your task is nothing else but computing parameters of some matrix ;) Here someone did this by hand, but I'd be surprised if there were no ready functions for that neither in the toolbox nor Matlab Central.

https://www.sancheya.com/computer-vision/camera-calibration-guide-matlab-code/

Good luck!

Przemysław,

Thank you for your help! I found a function in matlab. But I have only four points of one small object 10x6 cm. Is this enough - or do I have to live with the inaccuracy ? In https://www.sancheya.com/computer-vision/camera-calibration-guide-matlab-code/ there are a minimum 6 points required for the 6 extrinsic camera parameters.  

Thanks

Rupert

I think, the bigger the calibration sample the better accuracy. Hard to say if 10x6 cm is enough - best would be to compute the error of your measurements vs some known values.

Regarding intrinsic/extrinsic parameters, refer to the MathWorks link in my previous answer to figure that out. Perhaps you don't even need to perform the extrinsic transformation at all! AFAIK (though this is going slightly beyond my qualifications) this is needed if you want to precisely locate the objects you see in some real-world coordinates - useful for e.g. robotics. But if you only want to measure what you see, then I think it's sufficient to do the intrinsic transformation.