Skeleton fitting algorithm using energy minimization on geodesic distance features.

OpenSkeletonFitting (OpenSF) has been initiated as a research project at Osaka University, Japan from October 2011 to March 2012. It provides algorithms to enable skeleton tracking, using a single depth camera like Microsoft Kinect or Asus Xtion.

If you have ideas, constructive criticism or want to contribute to the project, please feel free to contact the author at anytime. Using this project or parts of it in derived works is highly encouraged. Please respect the licensing and mention the author "Norman Link norman.link@gmx.net".

You can find further information in the "Doc" directory. Please note that these files are a bit outdated, but they can still give a good overview of the project.

OpenSkeletonFitting (OpenSF) consists of the following modules:

is the main framework. It consists of all necessary utilities that are needed by the modules. This comprises module interface definitions, exception handling, logging and mathematical functions.

delivers the input data from a connected sensor. It can create input data from Microsoft Kinect, a .oni file and a numbered image stream.

retrieves the input frames from the sensor and segments it, such that the background is removed and the user can be detected. This currently only supports a simple background subtraction, in which the first captured frame is considered to show the static background and all the following frames are subtracted against this background. Only one single user can be detected.

uses the detected user silhouette and the corresponding depth pixels to reconstruct the point cloud and detect features that correspond to user end-affectors, i.e. hand, head, feet, torso. The end-affectors are detected by building a "geodesic distance map" of the detected user's point cloud, in which every pixel stores it's shortest distance from the user's center-of-mass. To be realtime-capable, a modified dijkstra's algorithm is used. The resulting distance map is then subsampled into "geodesic iso-patches", which are regions in the distance map of approximately the same distance from the center-of-mass. These iso-patches are then used to detect extrema in the distance map, which are assumed to correspond to end-affectors. "Geodesic lines" are then created by retrieving the geodesic path on the distance map from the end-affector to the center-of-mass.

This module also incorporates a simple feature tracking algorithm to create a trajectory for detected features and smooth the values using kalman filtering.

defines a skeleton hierarchy consisting of joints, bones and constraints. An energy function is then attached to each joint which uses the previously detected features, geodesic lines and the point cloud to define an energy, based on the joint parameters. The energy functions for the whole skeleton hierarchy are supposed to be minimal when the pose is optimal. A simple gradient minimization algorithm then uses the previous pose to change the skeleton parameters and minimize the overall skeleton energy.

A joint also defines classificator functions, which label the features to corresponding body parts, as well as extrapolator functions, which define the behaviour in case the body part has not been found.

is the main library to which other libraries can link against.

is a simple application which does not provide any visual output but only runs the algorithms.

visualizes the tracking results in either a 2d or 3d visualization.

The library has been successfully tested on Microsoft Windows 7, 8 and on MacOS X. For sake of simplicity, all required externals except boost are included in "Externals".

• Install boost 1.44 (you might as well try with later versions)
• define environment variable "BOOST_DIR" to point to the boost installation directory
• Open OpenSkeletonFitting with Visual Studio 2010
• build

• Install boost libraries
• Create project files using CMake from CMakeLists.txt
• build

This library is licensed under the GNU General Public License.