Action recognition based on parallel convolutional recurrent neural networks

To address the problem that a single feature of a video is insufficient for the representation of video motion information, we propose a parallel convolutional recurrent neural network-based action recognition. In this the parallel convolutional recurrent neural network, RGB image features and human joint point skeleton features were put into the CNN and RNN+LSTM, respectively. Furthermore, two features were connected into a joint spatio-temporal feature vector for finally action recognition. The experiment al results show that the action recognition accuracy of this paper method on UCF101 dataset is better than the current mainstream action recognition methods, which verifies the effectiveness of this method on action recognition.