Kinect for Windows: Find user height accurately

Vangos Pterneas blog

Kinect for Windows: Find user height accurately

  • Comments 6
  • Likes

Some people ask me if there is a way to determine the height of a user standing in front of a Kinect device. The answer is defintely positive: We can detect the user's height regardless of her distance from the sensor! We only need some very basic Maths knowledge. Let's find out more...

Download the sample demo application.

Prerequisites

The algorithm of finding the user's height

Kinect provides you with the coordinates (X, Y and Z) of 20 skeleton joints. Someone might think that a person's height is the distance from the head joint to a foot joint, right? Wrong! Your users might stand in bended knees, they might lean a bit to the left or to the right. If you try to calculate the distance defined from the head joint to one of the foot joints, you'll get a far from accurate result.

Another point of interest is that Kinect gives you the center of the head joint. This means that you'll need to add 9 - 12 centimetres to the calculated height. You won't be able to calculate the total height with 100% accuracy. If you need more accuracy, you'll need to detect the end of the head using the depth stream. Not that difficult, but let's focus on the skeleton stream right now.

If you examine the human skeleton joints a little more carefully, you'll notice that the height is the sum of the lengths of the following line segments:

  • Head - ShoulderCenter
  • ShoulderCenter - Spine
  • Spine - HipCenter
  • HipCenter - KneeLeft or KneeRight
  • KneeLeft / KneeRight - AnkleLeft / AnkleRight
  • AnkleLeft / AnkleRight - FootLeft / FootRight

Here is a comprehensive picture illustrating the above segments:

First of all, we need to calculate the length of the line defined by two joints in the 3D space. This is quite easy Mathematics. Simply find the square root of the sum of squared differences of the coordinates:

public static double Length(Joint p1, Joint p2)
{
    return Math.Sqrt(
        Math.Pow(p1.Position.X - p2.Position.X, 2) +
        Math.Pow(p1.Position.Y - p2.Position.Y, 2) +
        Math.Pow(p1.Position.Z - p2.Position.Z, 2));
}

And here is a way to find the length of more than two joints:

public static double Length(params Joint[] joints)
{
    double length = 0;

    for (int index = 0; index < joints.Length - 1; index++)
    {
        length += Length(joints[index], joints[index + 1]);
    }

    return length;
}

Pretty straightforward till here. But, wait a second! How will you determine which leg corresponds to the most accurate user height? Since we suppose that both of the user's legs have the same length, we need to choose the one that is tracked better! This means that no joint position is hypothesized. Here is how to find out the number of tracked joints within a joint collection:

public static int NumberOfTrackedJoints(params Joint[] joints)
{
    int trackedJoints = 0;

    foreach (var joint in joints)
    {
        if (joint.TrackingState == JointTrackingState.Tracked)
        {
            trackedJoints++;
        }
    }

    return trackedJoints;
}

We can now safely decide on which leg to use:

int legLeftTrackedJoints = NumberOfTrackedJoints(hipLeft, kneeLeft, ankleLeft, footLeft);
int legRightTrackedJoints = NumberOfTrackedJoints(hipRight, kneeRight, ankleRight, footRight);

double legLength = legLeftTrackedJoints > legRightTrackedJoints ? Length(hipLeft, kneeLeft, ankleLeft, footLeft) : Length(hipRight, kneeRight, ankleRight, footRight);

And here is the final method:

public static double Height(this Skeleton skeleton)
{
    const double HEAD_DIVERGENCE = 0.1;

    var head = skeleton.Joints[JointType.Head];
    var neck = skeleton.Joints[JointType.ShoulderCenter];
    var spine = skeleton.Joints[JointType.Spine];
    var waist = skeleton.Joints[JointType.HipCenter];
    var hipLeft = skeleton.Joints[JointType.HipLeft];
    var hipRight = skeleton.Joints[JointType.HipRight];
    var kneeLeft = skeleton.Joints[JointType.KneeLeft];
    var kneeRight = skeleton.Joints[JointType.KneeRight];
    var ankleLeft = skeleton.Joints[JointType.AnkleLeft];
    var ankleRight = skeleton.Joints[JointType.AnkleRight];
    var footLeft = skeleton.Joints[JointType.FootLeft];
    var footRight = skeleton.Joints[JointType.FootRight];

    // Find which leg is tracked more accurately.
    int legLeftTrackedJoints = NumberOfTrackedJoints(hipLeft, kneeLeft, ankleLeft, footLeft);
    int legRightTrackedJoints = NumberOfTrackedJoints(hipRight, kneeRight, ankleRight, footRight);

    double legLength = legLeftTrackedJoints > legRightTrackedJoints ? Length(hipLeft, kneeLeft, ankleLeft, footLeft) : Length(hipRight, kneeRight, ankleRight, footRight);

    return Length(head, neck, spine, waist) + legLength + HEAD_DIVERGENCE;
}

Finally, you can use this extension method within a custom application easily:

void Sensor_SkeletonFrameReady(object sender, SkeletonFrameReadyEventArgs e)
{
    using (var frame = e.OpenSkeletonFrame())
    {
        if (frame != null)
        {
            Skeleton[] skeletons = new Skeleton[frame.SkeletonArrayLength];
            frame.CopySkeletonDataTo(skeletons);

            var skeleton = skeletons.Where(s => s.TrackingState == SkeletonTrackingState.Tracked).FirstOrDefault();

            if (skeleton != null)
            {
                double height = skeleton.Height();
            }
        }
    }
}

You can now integrate this code to your own applications and find out the user's height! Download the complete WPF application in order to see it in action right now!

Enjoy kinecting!

Attachment: KinectUserHeight.zip
Your comment has been posted.   Close
Thank you, your comment requires moderation so it may take a while to appear.   Close
Leave a Comment
  • I am 1,97m, but the sample program tells me I am 1,72m. Any idea what could be wrong?

    Thanks!

  • @Christian Weyer

    This is a common problem tall people face. I think you are not totally visible from the depth cameras. Try to move further from the Kinect sensor and ensure that your head and feet state is "tracked" (not "hypothsized").

  • Had the same problem (though I am only 1.80) and I moved the sensor to "see" the floor and above..

  • Thanks for this sample. It was very helpful.

    The sample code seems to miss handling the very small amount of height contributed between hip center and either hip left or hip right. Perhaps this is intentional because it requires special handling to extract only the height component (i.e. the lateral distance from hip center to top of either of the hips is inappropriate). That contribution is likely so small as to be ignored for simplicity. Regardless, if I'm reading this code correctly, it diverges from both the introductory explanation and the skeleton graphic in two distinct ways: the intro paragraph skips hip left/right and the graphic includes the lateral distance between hip center and hip left/right.

  • Nice, but it is not enough!

  • I've used this and it's not accurate enough, even when using K4W v2 as well. Same issues, simply too inconsistent. One minute I'm 177cm, then 170, 180, even when standing still. Sampling over 3-5 seconds of frames etc. and bucketing data helps to throw bad frames, but still, it's not good enough.

Your comment has been posted.   Close
Thank you, your comment requires moderation so it may take a while to appear.   Close
Leave a Comment