This package contains modules that deal with gesture recognition using the Kinect (or Asus Xtion) sensor so as the stereo vision of the iCub. The gestureRecognition module implements a method to recognize gestures from a predefined set (examples of such gestures can be found in app/conf/supported_actions_easy.jpg or app/conf/supported_actions_hard.jpg). It is independent on the device and the operating system that is being used. Indeed, the KinectWrapper library, which is one of its dependencies, already provides an interface that deals with different devices and operating systems.

The module gestureRecognitionStereo instead, uses directly the stereo vision capabilities of the iCub robot. In this second case, gestures can be taught to the robot directly. Usually we train the first 5 gestures depicted in app/conf/supported_actions_easy.jpg.

The modules comprised here realize the memory game called All gestures you can. There are indeed two more modules, actionPerformer and gameManager, which have to be launched along with a LUA script.

• YARP
• icub-contrib-common
• OpenCV

For the game:

• speechRecognizer
• [iSpeak] (https://github.com/robotology/icub-main.git)
• LUA (Scripts written in LUA are employed to manage the speech recognition part of the demo.)

For gestureRecognition module using Kinect:

• kinect-wrapper

For gestureRecognitionStereo module using stereo vision:

• stereo-vision
• Hierarchical Image Representation

The project is composed of four modules. The gestureRecognition and gestureRecognitionStereo modules can be used independently from the others. If you don't have kinect, gestureRecognition will not be built. If you don't have stereo-vision, gestureRecognitionStereo will not be built.

In order to utilize this module, KinectServer (coded in kinect-wrapper) is to be running. Then, gestureRecognition can be launched. It will start immediately to recognize gestures belonging to the pool depicted in app/conf/supported_actions_easy.jpg.

It can run without any dependency. It requires some commands to start recognizing. In particular, you can write on the rpc port

• "pos" to save the initial position.
• "rec" to start recognizing.

Since the depth changes very much at every run, it is usually better to train the actions often. The database of the actions it is wanted to save and train is inside the context of the Hierarchical Image Representation folder, and it is called DBActions. In order to train new actions, it is possible to use the following commands using the LUA application:

• "save [one | two | three | four | five | six]"
• "let's train"

If you want to save more than six actions, it is necessary to slightly modify the LUA application file.

The game is structured in turns. One of the two players (the person or the robot) has to start performing a gesture. Then, the opponent recognizes the gesture, replicates it and adds a second random gesture to the sequence (from the pool of the possible gestures). At this point, the other player replicates the two gestures in the sequence, and adds a third gesture. This procedure is carried on until one of the two players loses.

The entire architecture can be run from the script files and can be seen in the following diagram:

Just launch the main application (app/scripts/demoGestureRecognition.xml if you are using kinect, app/scripts/demoGestureRecognitionStereo.xml if you are using stereo).

The commands that have to be provided are, respectively:

• "Let's start", to start the game. If the robot starts, it will say "It's your turn", and the person will just have to perform the gesture. Otherwise, the person will perform the gesture and then say "Your turn".
• "Your turn", as soon as the person is done performing the gestures.
• "You rock", after that the robot says "I think you are wrong", to let the robot know that it won.
• "You lost", after that the robot says "I think you are wrong", to let the robot know that it lost.
• "Game over", to quit the game straightaway.
• "Good bye", to close the LUA application.

The actions performed by the robot are specified in the file actions.ini if you are using kinect, and actionsStereo.ini if you are using stereo. In both cases it is easy to add or modify actions that you want the robot to perform.

Online documentation is available here: http://robotology.github.com/gesture-recognition.

Material included here is Copyright of Istituto Italiano di Tecnologia and EFAA Consortium. gesture-recognition is released under the terms of the GPL v2.0 or later. See the file LICENSE for details.