Semantic decoupling-based combined action recognition method for self-attention model

The invention discloses a combined action recognition method of a self-attention model based on semantic decoupling. The most advanced effect is achieved on three types of division of an STH-ELSE data set. According to the method, through an object-verb decoupling module (OMD) and a semantic decoupling constraint module (SDC), decoupling of advanced semantic features of verb and object combinations is realized, and the problem of model performance reduction caused by distribution deviation in combined action recognition is relieved. A plurality of initialized learnable marks are set in the OMD to capture spatio-temporal features related to an object, and the learned spatio-temporal features are preliminarily decoupled in a high-level visual space. And text information is introduced into the SDC to carry out stricter semantic-level consistency constraint on the decoupling features constructed in the OMD, and finally, completely decoupled appearance and motion features are learned. 本发明公开了一种基于语义解耦的自注意力模型的组合动作识别方