To obtain an action discriminating device which can learn fast and discriminate human's action by a simple operation processing.

A velocity vector is obtained from moving picture data, and a pattern vector is obtained by performing the Fourier expansion of the velocity vector. By giving the pattern vector to the input layer of a neural network which uses learning vector quantization, neurons corresponding to reference vectors which expresses respective classes ignite and the class which corresponds to a neuron which ignites most frequency is regarded as the discrimination result. The learning of the reference vector is performed by the learning vector quantization.