Työn laji — Arbetets art — Level Aika — Datum — Month and year Sivumäärä — Sidoantal — Number of pages

Gesture recognition is a hard task due to the presence of noise resulted from the unpredictability and ambiguity of the body motions. The joints ’ locations vary on all the axes which could add extra noise to the task of gesture recognition. Extra noise is added to the task as the target group of the research are small kids. On the other hand multiple gestures and similar features of some of them make the recognition task even harder, therefore multiple recognitions for different joints is needed to be done in parallel. Hidden Markov Models based techniques and the concept of threshold model are used to recognize the gesture motions from non-gesture motions. Firstly series of gestures are recorded and used to create the models. K-Means algorithm is used to cluster the points into the N states and labels the 3D points. Then the available alphabet of output symbols is expanded to M (M> N) states as it is not sure if the sequence of the points are a gesture or not. Next, by looking at the sequence of the labeled data it is possible to estimate how likely it is that the points have passed through the sequence the N states. Finally, if the likelihood is above the threshold a gesture is recognized