In May and June, the work was focused on the design and implementation of a movement gesture concept and corresponding task planning developments. Moreover an object recognition scheme was implemented for seamless gesture based object selection.              
Also, the ideas for the two stage gesture classification approach and the whole Human Robot Collaboration (HRC) system were presented at a workshop at the Robotics: Science and Systems conference in Berlin.

In March and April, the work was focused on the further development of the command gesture approach and the task planning module. Also, the ideas for the robust command gesture classification were submitted as a paper at a workshop at the Robotics: Science and Systems conference in Berlin.       
The previous command gesture approach was enhanced with a second processing stage, in order to deal with the high false positives rate of the gesture recognition.

In January and February, the work was focused on the development of probabilistic methods for the gesture based human robot interaction (tasks 4), and the task planning and task dependent adaption of robot motion planning (task 5). Also the deliverables 2.1 and 2.2 were completed.          
For the hand gesture analysis, a probabilistic hidden markov model based system was designed, developed, and integrated into the HRC framework.           
Also, a simple task planning module was designed and implemented, which processes the joint scene analyses of HMM-based action recognition and DL-based situation awareness.

In September and October 2012, the work was focused on gesture and action recognition. So far, it is based on Description Logics (DLs) which uses a taxonomy about actions, activities and gestures as a knowledge base. Recognition results are inferred directly by providing assertional knowledge which state information about human kinematics and robotic state. Moreover, preliminary linking concepts have been included into the knowledge base in order to allow for incorporation of stochastical methods for action and gesture recognition.

In May and June the work was focused on the training of the classifier for the body part detection, and the build-up of the multi-camera setup including the implementation of the calibration and registration step. Unfortunately the training of the classifier did not deliver the expected results, so that the existing approach will be used for the ICP tracking. After the installation of the sensors in the experiment setup, and the implementation of the calibration and registration step, the sensor system now consists of 3 Kinect and one time-of-flight sensor.