Self-supervised skeleton action recognition method based on cross-view consistency mining

The invention relates to the technical field of video processing, in particular to a self-supervised skeleton action recognition method based on cross-view consistency mining, which comprises the following steps of: acquiring and preprocessing label-free skeleton data, generating various views of a 3D skeleton from a skeleton sequence, and obtaining an amplification sequence of multi-view data in combination with various data enhancement methods. Encoding features of different views are obtained through an encoder, and a single-view comparison learning framework is established; further performing parallel self-supervised training on a plurality of view branches through contrast learning of instance discrimination to generate a plurality of single-view embedded features, constructing single-view semantic-level contrast learning through a nearest neighbor positive sample mining method, and learning collaborative representation of multiple views by combining a cross-view consistency mining module to obtain a mul