Action identification method, system and device based on optical flow guidance and double-flow network

The invention provides a pet dog action recognition method based on optical flow guide features and a double-flow network, and the method comprises the following steps: collecting video data containing pet dog actions, processing a video, and dividing the video into a training set and a test set; constructing a feature generation sub-network for extracting space appearance information of the video image sequence; building an OFF sub-network based on the optical flow guide features, wherein the OFF sub-network is used for extracting time motion information of the video image sequence; building a spatio-temporal information fusion network model, fusing the spatial feature map and the time feature map, and outputting a video-level feature vector; building an LK optical flow neural network based on the optical flow field; setting a model training hyper-parameter, training the model by using the training set, and storing a weight file obtained by training; and performing action prediction on the videos in the test