Action recognition method and device

The invention provides an action recognition method and device. The method comprises the following steps: acquiring multimedia video data including audio data and image data; analyzing the multimedia video data to obtain corresponding audio data and image data; extracting human body key point informatio from the image data, carrying out action recognition through a key point time sequence network, generating a first recognition result, whereinthe key point time sequence network is an action recognition network based on the key point information; converting the audio data into text information, analyzing the text information through a preset knowledge graph, and generating a second recognition result, wherein the preset knowledge graph is a knowledge graph with action names as nodes; and inputting the first identification result and the second identification result into a preset fusion network, and generating action prediction corresponding to the multimedia video data. Through fusion of the human body key poi