Emergency Rescue Action Recognition Algorithm Based on Recurrent -Adaptive Graph Convolutional Networks

Abstract

In the context of emergency rescue training, aiming at the scene of action recognition and effect evaluation by video. In order to solve the problems of large consumption of existing methods and poor real-time detection effect, this paper uses the technology stack of lightweight OpenPose and Recurrent-Adaptive Graph Convolutional Networks (R-AGCN). The recognition accuracy is improved by introducing cyclic enhancement module and adaptive module, and it has certain advantages for continuous video recognition. Firstly, the first ten layers of VGG - 19 network are used to extract image features, and OpenPose is used to extract bone feature key point coordinates. R-AGCN is used to realize action recognition. The module enhances the influence of spatial dimension on temporal dimension and then improve recognition accuracy. Under the two evaluation criteria of NTU-RGB + D data set, the accuracy of this algorithm is 84.3 % and 94.9 %, respectively, and it also has good recognition effect in the actual scene.