Human Action Recognition Based on Improved FCN Framework
Yixuan Cai, Hua Yu, Xuanzhe Fan, Yaqing Hou, Qiang Zhang · 2022
Human motion recognition is a highly active area of research. In this paper, we propose a Spatial Transformer Fully Convolutional Network (STFCN) for human action recognition, which leverages the advantages of both full convolution network (FCN) and spatial transformation network (STN). Firstly, in the stage of video image feature extraction, the proposed method integrates the STN network into the Convolutional Neural Networks (CNN) and the obtained feature maps are then passed through the FCN network. The upsampled operation of the FCN restores the feature maps to the size of the original input image. Finally, a softmax classifier is used to classify the human action at the pixel level. Extensive experiments are conducted on the standard human action datasets, i.e., HMDB51 and UCF101. The experimental results show that the STFCN achieves better performance than other compared methods.