Face Tracking in Kinect For Windows 1.5

After a long journey, my team at Microsoft shipped our real-time Face Tracking technology as face API in Kinect For Windows 1.5 in 2012! I worked on the 2D  & 3D face tracking technology (starting from the times when it was part of Avatar Kinect) and so I’d like to describe its capabilities and limitations in this post. For more information, see my paper “Real-time 3D face tracking based on active appearance model constrained by depth data”.

First of all, here is the face tracking demo:

You can use the Face Tracking SDK in your program if you install Kinect for Windows Developer Toolkit 1.5. After you install it, go to the provided samples and run/build yourself “Face Tracking Visualization” C++ sample or “Face Tracking Basics-WPF” C# sample. Off course, you need to have Kinect camera attached to your PC 😉 The face tracking engine tracks at the speed of 4-8 ms per frame depending on how powerful your PC is. It does its computations on CPU only (does not use GPU, since it may be needed to render graphics). For example, here is our face tracking used in Avatar Kinect app for XBox 360:

Below, you can see the code sample that uses our face API. It is relatively easy to add face tracking capabilities to your application. You need to link with a provided lib, place 2 dlls in the global path or in the working directory of your your executable (so they can be found) and add something like this to your code (this is in C++, you can also do it in C#, see the code samples):

// Include main Kinect SDK .h file
#include "NuiAPI.h"

// Include the face tracking SDK .h file
#include "FaceTrackLib.h"

// Create an instance of a face tracker
IFTFaceTracker* pFT = FTCreateFaceTracker();
if(!pFT)
{
    // Handle errors
}

// Initialize cameras configuration structures.
// IMPORTANT NOTE: resolutions and focal lengths must be accurate, since it affects tracking precision!
// It is better to use enums defined in NuiAPI.h

// Video camera config with width, height, focal length in pixels
// NUI_CAMERA_COLOR_NOMINAL_FOCAL_LENGTH_IN_PIXELS focal length is computed for 640x480 resolution
// If you use different resolutions, multiply this focal length by the scaling factor
FT_CAMERA_CONFIG videoCameraConfig = {640, 480, NUI_CAMERA_COLOR_NOMINAL_FOCAL_LENGTH_IN_PIXELS};

// Depth camera config with width, height, focal length in pixels
// NUI_CAMERA_COLOR_NOMINAL_FOCAL_LENGTH_IN_PIXELS focal length is computed for 320x240 resolution
// If you use different resolutions, multiply this focal length by the scaling factor
FT_CAMERA_CONFIG depthCameraConfig = {320, 240, NUI_CAMERA_DEPTH_NOMINAL_FOCAL_LENGTH_IN_PIXELS};

// Initialize the face tracker
HRESULT hr = pFT->Initialize(&videoCameraConfig, &depthCameraConfig, NULL, NULL);
if( FAILED(hr) )
{
    // Handle errors
}

// Create a face tracking result interface
IFTResult* pFTResult = NULL;
hr = pFT->CreateFTResult(&pFTResult);
if(FAILED(hr))
{
    // Handle errors
}

// Prepare image interfaces that hold RGB and depth data
IFTImage* pColorFrame = FTCreateImage();
IFTImage* pDepthFrame = FTCreateImage();
if(!pColorFrame || !pDepthFrame)
{
    // Handle errors
}

// Attach created interfaces to the RGB and depth buffers that are filled with
// corresponding RGB and depth frame data from Kinect cameras
pColorFrame->Attach(640, 480, colorCameraFrameBuffer, FORMAT_UINT8_R8G8B8, 640*3);
pDepthFrame->Attach(320, 240, depthCameraFrameBuffer, FTIMAGEFORMAT_UINT16_D13P3, 320*2);
// You can also use Allocate() method in which case IFTImage interfaces own their memory.
// In this case use CopyTo() method to copy buffers

FT_SENSOR_DATA sensorData;
sensorData.pVideoFrame = &colorFrame;
sensorData.pDepthFrame = &depthFrame;
sensorData.ZoomFactor = 1.0f;       // Not used must be 1.0
sensorData.ViewOffset = POINT(0,0); // Not used must be (0,0)

bool isFaceTracked = false;

// Track a face
while ( true )
{
    // Call Kinect API to fill videoCameraFrameBuffer and depthFrameBuffer with RGB and depth data
    ProcessKinectIO();

    // Check if we are already tracking a face
    if(!isFaceTracked)
    {
        // Initiate face tracking.
        // This call is more expensive and searches the input frame for a face.
        hr = pFT->StartTracking(&sensorData, NULL, NULL, pFTResult);
        if(SUCCEEDED(hr) && SUCCEEDED(pFTResult->Status))
        {
            isFaceTracked = true;
        }
        else
        {
            // No faces found
            isFaceTracked = false;
        }
    }
    else
    {
        // Continue tracking. It uses a previously known face position.
        // This call is less expensive than StartTracking()
        hr = pFT->ContinueTracking(&sensorData, NULL, pFTResult);
        if(FAILED(hr) || FAILED (pFTResult->Status))
        {
            // Lost the face
            isFaceTracked = false;
        }
    }

    // Do something with pFTResult like visualize the mask, drive your 3D avatar,
    // recognize facial expressions
}

// Clean up
pFTResult->Release();
pColorFrame->Release();
pDepthFrame->Release();
pFT->Release();

The code calls the face tracker by using either StartTracking() or ContinueTracking() functions. StartTracking() is a more expensive function since it searches for a face on a passed RGB frame. ContinueTracking() method uses previous face location to resume tracking. StartTracking() is more stable when you have big breaks between frames since it is stateless.

There are 2 modes in which the face tracker operates – with skeleton based information and without. In the 1st mode you pass an array with 2 head points to StartTracking/ContinueTracking methods. These head points are the end of the head bone contained in NUI_SKELETON_DATA structure returned by Kinect API. This head bone is indexed by NUI_SKELETON_POSITION_HEAD member of NUI_SKELETON_POSITION_INDEX enumeration. The 1st head point is the neck position and the 2nd head point is the head position. These points allow the face tracker to find a face faster and easier, so this mode is cheaper in terms of computer resources (and sometimes more reliable at big head rotations). The 2nd mode only requires color frame + depth frame to be passed with an optional region of interest parameter that tells the face tracker where to search on RGB frame for a user face. If the region of interest is not passed (passed as NULL), then the face tracker will try to find a face on a full RGB frame which is the slowest mode of operation of StartTracking() method. ContinueTracking() will use a previously found face and so is much faster.

Camera configuration structure –  it is very important to pass correct parameters in it like frame width, height and the corresponding camera focal length in pixels. We don’t read these automatically from Kinect camera to give more advanced users more flexibility. If don’t initialize them to the correct values (that can be read from Kinect APIs), the tracking accuracy will suffer or the tracking will fail entirely.

Frame of reference for 3D results –  the face tracking SDK uses both depth and color data, so we had to pick which camera space (video or depth) to use to compute 3D tracking results in. Due to some technical advantages we decided to do it in the color camera space. So the resulting frame of reference for 3D face tracking results is the video camera space. It is a right handed system with Z axis pointing towards a tracked person and Y pointing UP. The measurement units are meters. So it is very similar to Kinect’s skeleton coordinate frame with the exception of the origin and its optical axis orientation (the skeleton frame of reference is in the depth camera space). Online documentation has a sample that describes how to convert from color camera space to depth camera space.

Also, here are several things that will affect tracking accuracy:

1) Light – a face should be well lit without too many harsh shadows on it. Bright backlight or sidelight may make tracking worse.

2) Distance to the Kinect camera – the closer you are to the camera the better it will track. The tracking quality is best when you are closer than 1.5 meters (4.9 feet) to the camera. At closer range Kinect’s depth data is more precise and so the face tracking engine can compute face 3D points more accurately.

3) Occlusions – if you have thick glasses or Lincoln like beard, you may have issues with the face tracking. This is still an open area for improvement 🙂  Face color is NOT an issue as can be seen on this video

Here are some technical details for more technologically/math minded people: We used the Active Apperance Model as the foundation for our 2D feature tracker. Then we extended our computation engine to use Kinect’s depth data, so it can track faces/heads in 3D. This made it much more robust compared to 2D feature point trackers. Active Appearance Model is not quite robust to handle all real world scenarios. Off course, we also used lots of secret sauce to make things working well together 🙂  You can read about some of these algorithms here, here and here.

Have fun with the face tracking SDK!

We published this paper that describes the face tracking algorithm in details.

Credits – Many people worked on this project or helped with their expertise:

Christian Hutema (led the project), Lin Liang, Nikolai Smolyanskiy, Sean Anderson, Evgeny Salnikov, Jayman Dalal, Jian Sun, Xin Tong, Zhengyou Zhang, Cha Zhang, Simon Baker, Qin Cai.