Human behavior recognition based on attention mechanism

In recent years, human behavior recognition has become a research hotspot in the field of computer vision. Compared with static images, human behavior recognition is based on video analysis. The video contains two parts of information in the spatial and temporal domains. The spatiotemporal two-stream convolutional neural network model gives the same weight to different spatial and temporal cues in the parameter training stage, which seriously affects the distinction of features. In response to this shortcoming, this paper proposes to introduce attention mechanism in the spatiotemporal dual-flow network model. This method can effectively add a weight to the static image features and the dense optical flow features between frames, pay attention to the beneficial regions in the feature information and then distinguish the feature information, so as to achieve more accurate behavior recognition. Experiments were conducted on the UCF-101 human behavior video dataset, and the results prove the effectiveness of the spatiotemporal two-stream convolutional neural network with attention mechanism introduced in the human behavior recognition task.