Self-supervised action recognition method and device based on hierarchical multi-view

The invention discloses a self-supervised action recognition method and device based on hierarchical multi-view, and the method comprises the steps: firstly, in a pre-training stage of comparative learning, original data is subjected to data enhancement and recombination, and features are extracted through a key encoder and a query encoder respectively to construct positive and negative samples; during feature extraction, sub-view angles of different scales are obtained through a view angle generation module, and time features, space features and combinations of the time features and the space features are obtained through Transform coding, aggregation and splicing. The spatio-temporal features achieve the targets of approaching a positive sample and pushing away a negative sample through loss comparison so as to obtain the feature representation capability. And the feature distance between different visual angles is shortened through the consistency loss between the time feature and the space feature. The me