Methods, systems, media, and apparatus for unsupervised action detection

The invention discloses a method, a system, a medium and equipment for unsupervised action detection, and belongs to the field of computer vision. According to the method, firstly, a skeleton encoder capable of extracting action semantics is trained in an unsupervised training mode; and training a linear classifier on a small number of labeled skeleton sequences, and obtaining an excellent action detection system by using the trained skeleton encoder and the linear classifier. An iterative training process is designed to ensure that an unsupervised training skeleton encoder can obtain more accurate semantic information, and the method mainly comprises two stages: 1) comparative learning: inputting a skeleton sequence, generating a positive sample through data enhancement, and generating a negative sample from a memory bank; 2) semantic clustering: encoding all skeleton sequences, mining action semantic information through clustering, outputting a current semantic code book, and optimizing a next round of comp