Gesture recognition engines rely on machine learning techniques to learn and detect gestures. Algorithms such as Dynamic Time Warping [3, 4], Linear Time Warping [4], One Dollar recogniser [2], Rubine (3D) [1], SiGeR [1], Support Vector Machines [3], Neural Networks [3], and Hidden Markov Models (HMM) [5, 6] are well known in this domain. Although each recogniser claims to be a complete solution for gesture detection problems, we experienced that some algorithms either provide false positives very easily while others are more reserved. (For example: open http://depts.washington.edu/aimgroup/proj/dollar/ and try to draw a clockwise circle).

In this thesis, we plan to fuse different machine learning techniques to improve their gesture recognition rate. This means we are going to use two or three different recognisers at the same data and describe in what situation we can achieve a higher fidelity ratio.

1. Port existing recognisers to Midas (programmed in C)
2. Find a way to express the fusion mechanism (domain-specific language, ...)
3. Perform an evaluation and show which techniques work best depending on the properties of the gesture and what combination of algorithms generally provides the best results.