Recognizing Human Actions By Fusing Spatio-temporal Appearance And Motion Descriptors