gloveTrack

gloveTrack is a free, libre and open source hand tracking system intended for (eventual) use with Virtual Reality headsets. Currently key portions are implemented, but the important functionality such as full lookup (and non-manual image normalisation training) is still unimplemented, limiting its applications.

gloveTrack intends to let your hands enter computer games (and other 3D rendered environments) to interact with objects, requiring only common household items for this to occur: a camera, latex gloves and some colored whiteboard markers.

gloveTrack is an implementation of most of the techniques described in Robert Y. Yang's 2009 research paper "Real-Time Hand-Tracking with a Color Glove".

Contributions are highly welcome.

TODO

gloveTrack is developed as a side-project for fun, with minimal deadlines to help further develop personal software development skills. As gloveTrack matures with reliabilty and dependability becoming the top priority, more rigor will be placed on development (rather being more of a sandbox for OpenCV experiments and prototypes).

Here are some key items on gloveTrack the TODO list:

[ ] Write accelerated search algorithm with unit tests (see research paper)

[ ] Generate set of (100,000) handpose images covering the human handpose space (currently testing with 100 image subset)

[ ] Render the generated handpose estimate using LibHand

[ ] Polish command line arguments (and replace getopt() with Boost)

[ ] Automated training of image normalization model (stretch goal)

Related to the development of gloveTrack is the enhancement of Marin Saric's LibHand:

[ ] Enhance LibHand to allow rendering at arbitrary x, y co-ordinates (currently limited to camera on surface of sphere)

[ ] Move away from OGRE3D to more suitable cross-platform rendering library

Physical dependencies

• Video source with accurate color representation (most webcams should be fine, or smartphone connected as USB camera)
• Colored gloves
• Moderately powerful PC (future optimization will allow lower spec computers such as smartphones to run gloveTrack)

Sofware dependencies

gloveTrack was developed against software dependencies

• C++11 compiler
• OpenCV (Open Source Computer Vision) v2.4.9
• CMake v2.8 (v3.4 on Windows)
• LibHand v0.9 by Marin Saric (supplied as git submodule), which was initially developed against
  • OpenCV (Open Source Computer Vision) v2.3.y
  • OGRE (Object Orientated Graphics Rendering Engine) v1.7.4
    • OpenGL headers (GL/glu.h)
  • Tcl/Tk (likely v8.5.y)
  • CMake v2.8.6
• Git Large File Storage (LFS) v1.5.5

Supported Platforms

While gloveTrack and dependencies are cross-platform, development takes place on Debian so satisfying the dependencies on Ubuntu-derived Linux distributions is just apt-get away. Unfortunately LibHand's OGRE dependency makes satisfying dependencies on macOS and Windows much more time consuming (see LibHand's README.md for more details), so it's not worth pursuing those platforms at this stage. Future versions of LibHand will allow gloveTrack to expands its officially supported platforms.

Satisfying software dependencies under Debian/Ubuntu

sudo apt-get install build-essential cmake git libogre-dev libglu1-mesa-dev libxt-dev libopencv-dev tcl tk

On Ubuntu 16.04, substitute libogre-dev with libogre-1.9-dev.

Additionally, Git LFS must be installed to retrieve the developmental image and video sets from the 'db' submodule.

Compiling:

git clone https://github.com/[user]/gloveTrack.git
cd gloveTrack
git submodule init
# Download LibHand submodule and 'db' submodule
git submodule update --recursive
# Download db submoudle's ~16MB submodule containing image/video sets via Git LFS
cd db
git lfs fetch
# Git LFS references don't update to large files without this
git lfs pull
cd ..

# Build LibHand (see https://github.com/libhand/libhand for detailed installation instructions) 
cd third-party/libhand/ && mkdir build && cd build/
cmake ..
make
# Collate build artifacts in *local* directory
make install

# Build gloveTrack
cd ../../../
mkdir build
cd build

cmake ..
make

Example usage:

All example from root gloveTrack directory.

# Use saved video as input and enable generation rendering of 3D model search set
./build/src/gloveTrack --display-width=400 --display-height=400 --processing-width=20 --processing-height=20 --pre-crop-width 200 --pre-crop-height 200 --input-video db/evaluationSet/fullFrameVideo.mp4 --generate-search-set

# Use camera as input
./build/src/gloveTrack --display-width=200 --display-height=200 --processing-width=30 --processing-height=30

# Use saved video as input
 ./build/src/gloveTrack --display-width=400 --display-height=400 --processing-width=20 --processing-height=20 --pre-crop-width 200 --pre-crop-height 200 --input-video db/evaluationSet/vid.mp4