A Human Action Recognition Algorithm in Dynamic Scene of Emergency Rescue

Abstract

Action recognition is a hot topic in computer vision. As an emerging interdisciplinary discipline, rescue medicine has been paid more and more attention by more countries and industries. However, there is currently fewer research on action recognition in emergency rescue scenes, and this problem is susceptible to dynamically changing noisy background and other factors. Aiming at the problem, this paper makes the dataset and proposes a GC-SlowFast network combining the attention mechanism. The proposed algorithm combines the SlowFast network with the attention module global context block (GC Block), which can effectively integrate the dependent information between remote key features on the basis of the original network, weaken the interference of the background on target action recognition, and improve the accuracy of action recognition. With R(2+1)D as the backbone of GC-SlowFast, the number of parameters can be significantly reduced by optimizing the network structure. In addition, in view of the problem for small data easily leading to overfitting, this paper enhances data on the self-built dataset. GC-SlowFast is pre-trained on the HACS Clips dataset and then fine-tuned on the self-built dataset.