Gesture recognition based on SOM using multiple sensors

Gesture recognition is important, because it is a useful communication medium between humans and computers. In this paper we use multiple sensors, i.e., PSD cameras for detecting LEDs on a body and DataGloves for both hands. One of the major difficulties in gesture recognition is temporal segmentation from continuous motion. We use training samples with manually attached labels as prior knowledge. A self-organizing map(SOM) is constructed based on training samples. Test gestural data is segmented by systematic search to obtain the best match with a reference vector in the resulting SOM. A comparative study is done between the use of a single SOM and 3 SOMs for representing spatio-temporal information obtained from PSD cameras and DataGloves.