Ask Your Question

How to calculate the head and body angle with respect to the camera?

asked 2019-01-15 21:53:39 -0500

prb gravatar image

updated 2019-01-23 04:02:32 -0500

I would like to calculate the human body angle with respect to the camera. I have attached a screenshot for your kind reference. If the person is looking into the camera then the angle is zero.
If the person is looking up then the angle is more than 0. I have the pose coordinates of the person. I would appreciate your thoughts on this. Thank you. C:\fakepath\Screen Shot 2019-01-16 at 12.38.55 PM.png

EDIT: Duplicate Question

edit retag flag offensive close merge delete



maybe use dlib to get face landmark then caculate the angle like:

gino0717 gravatar imagegino0717 ( 2019-01-15 23:01:29 -0500 )edit

i think, you need a "head pose estimation", maybe something like this

berak gravatar imageberak ( 2019-01-16 06:54:49 -0500 )edit

I used this tensorflow code for head pose estimation a while back and it worked well for me

phillity gravatar imagephillity ( 2019-01-26 13:55:14 -0500 )edit

2 answers

Sort by ยป oldest newest most voted

answered 2019-01-17 02:57:36 -0500

berak gravatar image

updated 2019-01-18 03:35:24 -0500

i still think, you want a head pose estimation here. the gist is: you can use solvePnP to get an rvec and a tvec, containing rotation / translation wrt. the camera.

for this, you will need 2d landmarks (e.g from dlib, openpose, or from opencv's landmark models) and corresponding 3d points. i'll try with FacemarkLBF here:

// create and load the facemarks model
cv::Ptr<cv::face::Facemark> facemark;
facemark = cv::face::createFacemarkLBF();

// load the 68 3d points from file (see below !)
std::vector<cv::Point3d> pts3d;
cv::FileStorage fs2("points3d.yml",0);
fs2["points"] >> pts3d;

then for each image, detect a face, then get the current landmarks:

std::vector<cv::Rect> rects;
face_cascade.detectMultiScale(gray_img, faces, 1.4, 2, cv::CASCADE_SCALE_IMAGE, cv::Size(30, 30));

std::vector<cv::Rect> faces(1,rects[0]);
std::vector< std::vector<cv::Point2f> > shapes;

std::vector<cv::Point2d> &pts2d;
for(size_t k=0; k<shapes[0].size(); k++)

now we can apply solvePnP:

// if you did not calibrate it, use a camMatrix based on img size:
cv::Mat rvec,tvec;
cv::Mat camMatrix;
int max_d = std::max(s.width,s.height);
camMatrix = (cv::Mat_<double>(3,3) <<
    max_d,   0, s.width/2.0,
    0,     max_d, s.height/2.0,
    0,   0,        1.0);

// 2d -> 3d correspondence
cv::solvePnP(pts3d, pts2d, camMatrix, cv::Mat(1,4,CV_64F,0.0), rvec, tvec, false, cv::SOLVEPNP_EPNP);

image description

the whole code is here

and, for your convenience, here's points3d.yml:

points: [ -7.4077796936035156e+001, 4.5610500335693359e+001,
    1.7611330032348633e+001, -7.6078399658203125e+001,
    2.4455335617065430e+001, 1.4652364253997803e+000,
    -7.0680282592773438e+001, 3.8770267963409424e+000,
    1.6104341506958008e+001, -6.9542381286621094e+001,
    -1.8663349151611328e+001, -5.0522098541259766e+000,
    -6.0891132354736328e+001, -3.7201663970947266e+001,
    -4.9577393531799316e+000, -4.7551403045654297e+001,
    -5.1671474456787109e+001, 2.1515935897827148e+001,
    -3.3833507537841797e+001, -6.4209655761718750e+001,
    3.3763854980468750e+001, -1.8493196487426758e+001,
    -7.5527656555175781e+001, 4.2787197113037109e+001,
    -2.5200850963592529e+000, -7.7372253417968750e+001,
    4.5473331451416016e+001, 1.3970505714416504e+001,
    -7.3213897705078125e+001, 6.4313529968261719e+001,
    3.0962856292724609e+001, -6.6279350280761719e+001,
    3.5533737182617188e+001, 4.6108547210693359e+001,
    -5.3055961608886719e+001, 1.9751256942749023e+001,
    5.1060363769531250e+001, -3.2454692840576172e+001,
    6.6386039733886719e+001, 5.8377300262451172e+001,
    -1.4232730865478516e+001, 5.9445739746093750e+001,
    6.3227752685546875e+001, 4.3665418624877930e+000,
    4.6228359222412109e+001, 7.0979812622070312e+001,
    2.6926740646362305e+001, 4.6090355515480042e-001,
    6.5315643310546875e+001, 4.6220058441162109e+001,
    4.7723823547363281e+001, -5.8375602722167969e+001,
    5.7522071838378906e+001, 7.4911415100097656e+001,
    -4.7820030212402344e+001, 6.2965705871582031e+001,
    9.6279922485351562e+001, -3.4796894073486328e+001,
    6.4059089660644531e+001, 1.0583456420898438e+002,
    -2.3467020034790039e+001, 6.1613960266113281e+001,
    1.1014395904541016e+002, -1.2404515266418457e+001,
    5.8514854431152344e+001, 1.1110581970214844e+002,
    9.6587600708007812e+000, 5.9617969512939453e+001,
    1.1002123260498047e+002, 2.0711694717407227e+001,
    6.3654747009277344e+001, 1.0981579589843750e+002,
    3.2074752807617188e+001, 6.5145515441894531e+001,
    1.0512077331542969e+002, 4.5245258331298828e+001,
    6.3173934936523438e+001, 9.4144226074218750e+001,
    5.4559543609619141e+001, 5.6469257354736328e+001,
    7.4634750366210938e+001, -1 ...
edit flag offensive delete link more


HMMmmmnnn. No python. Too much work for me.

supra56 gravatar imagesupra56 ( 2019-01-17 06:50:00 -0500 )edit

@supra56, again look at the learnopencv site, it has python code for the same idea, too.

berak gravatar imageberak ( 2019-01-17 06:57:10 -0500 )edit

@berak. I know that link.

supra56 gravatar imagesupra56 ( 2019-01-17 07:08:41 -0500 )edit

wow wait, you got the complete solution, and then you simply say, ah no time to work on it, screw this xD

StevenPuttemans gravatar imageStevenPuttemans ( 2019-01-18 03:12:03 -0500 )edit

answered 2019-01-16 04:43:10 -0500

updated 2019-01-20 10:44:23 -0500

berak gravatar image

Like stated above, you will need more than the basic functionality provided by OpenCV. Look at gaze estimation algorithms. I know of a paper that uses Cascade Classifiers to define a rough angle, and which is fully integrated into OpenCV. Have a look here:

@StevenPuttemans, hope you don't mind me "pirating" your answer, but given it's so simple, -- it works amazingly well !! ;)

VideoCapture cap(0);
CascadeClassifier profile("c:/p/opencv/data/haarcascades/haarcascade_profileface.xml");
while(1) {
    Mat frame, gray;;
    cv::cvtColor(frame, gray, cv::COLOR_BGR2GRAY);
    std::vector<cv::Rect> faces_right,faces_left;
    std::vector<int> lvl_right,lvl_left;
    std::vector<double> weights_right,weights_left;
    // right face side
    profile.detectMultiScale(gray, faces_right, lvl_right, weights_right, 1.2, 1, 0, cv::Size(30, 30), Size(), true);
    // flip, and apply again for left one
    profile.detectMultiScale(gray, faces_left, lvl_left, weights_left, 1.2, 1, 0, cv::Size(30, 30), Size(), true);
    float angle = 0; // formula from paper: a=-90*l+90*r ;)
    if (weights_right.size()>0 && weights_right[0]>0)
        // 4: heuristic scale factor, though i'm pretty sure, this is not linear ;)
        angle += 90 * weights_right[0] / 4; 
    if (weights_left.size() && weights_left[0]>0)
        angle += -90 * weights_left[0] / 4;
    cout << angle << endl;

    if (waitKey(10) > -1) break;
edit flag offensive delete link more


Thanks, Steven. I forgot to mention that I have pose coordinates. Is finding the center of the head and center of camera enough? By that, I will find the distance between the head and the camera. After that, I can calculate the angle between this line and the of the body parts. Does that make sense?

prb gravatar imageprb ( 2019-01-16 05:08:10 -0500 )edit

there some issues

  • you will never get the physical distance to a camera, unless you have a reference size somewhere
  • people can rotate their head unrelated to their body parts
StevenPuttemans gravatar imageStevenPuttemans ( 2019-01-16 06:19:55 -0500 )edit

I forgot to mention that I have pose coordinates.

what exactly do you have ? you will need the rotation of the head (wrt the camera), the position is useless (imho)

berak gravatar imageberak ( 2019-01-16 06:26:05 -0500 )edit

@StevenPuttemans nice link, but it only approximates rotation around y

berak gravatar imageberak ( 2019-01-16 06:29:35 -0500 )edit

@StevenPuttemans . Does my measuring angle will work in front of camera? The answer is YES. Can do both horizontal or vertical or both. Check this on youtube watch?v=E5OIhlCYGyw.

supra56 gravatar imagesupra56 ( 2019-01-16 06:54:11 -0500 )edit

@berak, thats because we did not take into account the X axis indeed :D @supra56 for measuring around the X axis you will need two other models trained first for this to work, might be overkill.

Also in general, if you have an angle dataset, you can let a DL network predict the angle directly. As long as you have annotated data. Did you consider for example OpenPose?

StevenPuttemans gravatar imageStevenPuttemans ( 2019-01-16 08:49:58 -0500 )edit

You guys are awsome.

holger gravatar imageholger ( 2019-01-16 16:37:17 -0500 )edit

@StevenPuttemans -- is the code for the paper you mention still available ? (link to is dead)

@prb if you have the openpose face model, you could derive rotation around z without any further means, but imho, for anything else you'd still need solvePnP() (and 3d reference points)

berak gravatar imageberak ( 2019-01-18 01:59:07 -0500 )edit

@berak I went through the same code in python on ( But I still not able to understand how I will be able to find the head angle. As mentioned in the comment section of the website, the author was talking about using eulerAnglesToRotationMatrix : github link

prb gravatar imageprb ( 2019-01-18 02:23:00 -0500 )edit

@prb again, the rvec from solvePnP will contain the rotation angles (in radians) towards the camera. how to process those further depends on your application

berak gravatar imageberak ( 2019-01-18 02:30:34 -0500 )edit

Question Tools


Asked: 2019-01-15 21:53:39 -0500

Seen: 4,305 times

Last updated: Jan 23 '19