Gesture recognition using accelerometers and machine learning

Human gesture recognition is the ability of a machine to recognize human gestures. It is used in various fields such as security, sports, exercise, medicine, robotics, computer interfaces, virtual reality, games… There are many different ways of obtaining human motion data. Recently we have seen increased usage of accelerometers in mobile phones and consumer electronics for this purpose. Reasons for that can be found in advances in MEMS technology which resulted in accelerometers that are as small as 3mm x 5mm x 0.9mm, operate on less than 1 milliwatt of power, and cost less than one dollar. Despite the explosion of accelerometers usage in consumer electronics, most of the gesture recognition applications are still quite basic (window rotation) or very problem specific. This indicates that there is a lack of general approach to gesture recognition using accelerometers that would enable training of the gestures. For this purpose two machine learning methods have been examined in this project: k-nearest neighbors and decision tree induction from time series. Both methods were evaluated on two gesture recognition domains: handwritten letter recognition (letters were written on the table using two-axis accelerometer) and hand signals (signals were given by moving three-axis accelerometer in the air). The methods were evaluated on data without noise as well on data with added noise. Experimental results have shown that both methods perform very well on given domains of gesture recognition