In order to avoid the influence of external factors on the subsequent recognition of RGB video and improve the accuracy of human motion recognition, an algorithm of human action recognition based on Two-Stream Ind Recurrent Neural Network is proposed. In terms of extracting features, the temporal network extracts the information on the 3D coordinate of different joints at each time and classifies it by a softmax layer. The spatial network converts the spatial positional relationship of the joints at each moment into a skeleton sequence and inputs it into the softmax layer to classify. Finally, the results of the classification of the temporal network and the spatial network are weighted and summed to obtain the final classification result. Experiments verify the validity of the model on the largest 3D skeleton action recognition dataset NTU RGB + D and SBU interactive dataset.
Paper
Full text
Human action recognition based on two-stream Ind recurrent neural network
Semantic Scholar · Computer Science · 2019
Abstract
In order to avoid the influence of external factors on the subsequent recognition of RGB video and improve the accuracy of human motion recognition, an algorithm of human action recognition based on Two-Stream Ind Recurrent Neural Network is proposed. In terms of extracting features, the temporal network extracts the information on the 3D coordinate of different joints at each time and classifies it by a softmax layer. The spatial network converts the spatial positional relationship of the joints at each moment into a skeleton sequence and inputs it into the softmax layer to classify. Finally, the results of the classification of the temporal network and the spatial network are weighted and summed to obtain the final classification result. Experiments verify the validity of the model on the largest 3D skeleton action recognition dataset NTU RGB + D and SBU interactive dataset.