gazr is a library and a set of tools for real-time face tracking and gaze estimation from a monocular camera (typically, a webcam) or a RGB-D stream (3D camera).
It natively supports ROS.
Currently, it only performs 6D head pose estimation. Eye orientation based on pupil tracking is being worked on.
If you plan to use this library for academic purposes, we kindly request you to cite our work.
This library (libhead_pose_estimation.so) performs 3D head pose estimation
based on the fantastic dlib face detector and a bit of
OpenCV's
solvePnP magic (it uses adult male anthropometric data to match a real 3D head to the projected image).
The library returns a 4x4 transformation matrix.
It supports detection and tracking of multiple faces at the same time, and runs on-line, but it does not provide face identification/recognition.
If provided with an RGB-D (color + depth) stream, the library can extract and compute the 3D localisation of 68 facial landmarks (cf screenshot above).
Note that this feature is currently only available for ROS.
Note: The library has only been tested on Linux. We can only provide limited support for other operating systems!
You need to download and extract Dlib somewhere. This
application requires dlib >= 18.18. On Ubuntu 16.04 and above, sudo apt-get install libdlib-dev
You also need OpenCV. On Ubuntu, sudo apt-get install libopencv-dev. The code
has been last tested with OpenCV4. Previous versions of OpenCV might or might
not work.
The library uses a standard CMake workflow:
$ mkdir build && cd build
$ cmake -DCMAKE_BUILD_TYPE=Release ..
$ make
Do not forget to enable Release mode, otherwise the performances will be terrible.
You can test the result with:
$ ./gazr_show_head_pose ../share/shape_predictor_68_face_landmarks.dat
Finally, if desired, you can install the library and the executables:
$ make install
By default, only the sample tools are built (CMake parameter WITH_TOOLS=ON).
See below for ROS support.
The ROS wrapper provides a convenient node that exposes each detected face as a TF frame.
Before building gazr with the ROS wrapper, make sure that you have installed the following ROS- < distro > package dependencies, where < distro > is the ROS distribution in your machine.
The code has been tested up to ROS noetic, but should work for any ROS1 distribution.
For example, this is the case for ROS-kinetic distribution:
sudo apt-get install ros-kinetic-roscpp ros-kinetic-tf ros-kinetic-std-msgs ros-kinetic-visualization-msgs ros-kinetic-sensor-msgs ros-kinetic-cv-bridge ros-kinetic-image-transport ros-kinetic-image-geometry
The compilation of the ROS wrapper is disabled by default. You can enable it with:
cmake -DWITH_ROS=ON
Once installed with ROS support, you can launch gazr with a monocular RGB
stream with:
$ roslaunch gazr gazr.launch
The estimated TF frames of the heads will then be broadcasted as soon as detected.
The number of detected faces is published on /gazr/detected_faces/count and if
gazr has been compiled with the flag DEBUG_OUTPUT=TRUE, then the detected
features can be seen on the topic /gazr/detected_faces/image.
To process a depth stream as well, run:
$ roslaunch gazr gazr.launch with_depth:=true
The facial features are published as a PointCloud2 message on the
/gazr/facial_features topic.
You can get the full list of arguments by typing:
$ roslaunch gazr gazr.launch --ros-args
Importantly, you might want to remap the rgb and depth topics to your liking.
Run ./gazr_show_head_pose ../share/shape_predictor_68_face_landmarks.dat to test
the library. You should get something very similar to the picture above.
Run ./gazr_estimate_head_pose ../share/shape_predictor_68_face_landmarks.dat frame.jpg
to print the head pose detected in frame.jpg.
Run ./gazr_estimate_head_pose ../share/shape_predictor_68_face_landmarks.dat image_file_names.txt
to print the head pose detected in each image file listed in image_file_names.txt (image file names written in new lines).