Classification of social gestures : Recognizing waving using supervised machinelearning

Abstract: This paper presents an approach to gesture recognition including the use of a tool in order to extract certain key-points of the human body in each frame, and then processing this data and extracting features from this. The gestures recognized were two-handed waving and clapping. The features used were the maximum co-variance from a sine-fit to time-series of arm angles, as well as the max and min of this fitted sinus function. A support vector machine was used for the learning. The result was a promising accuracy of 93% ± 4% using 5-fold cross-validation. The limitations of the methods used are then discussed, which includes lack of support for more than one gesture in the data as well as some lack of generality in means of the features used. Finally some suggestions are made as to what improvements and further explorations could be made.