Multiscale Graph Convolutional Networks-based Action Recognition and motion Prediction

To address the challenges of human behavior prediction, we propose a multi-scale graph convolutional network based on the improvement of traditional graph convolutional networks. And by fusing the prediction task and the recognition task, the features extracted at each stage of the network inferencing process are shared to assist in improving the prediction performance of the prediction network. Our goal is to explore a new method can accurately extracting human action sequences in human behavior prediction tasks. We have constructed a complete network framework based on our proposed method. And we have done the validation of our proposed method, we use Human3.6M dataset and 3DPW dataset as our training data, the prediction error of our method on H3.6M dataset is optimal compared to the previous methods, although the short-term prediction on 3DPW dataset is not better than previous methods, but in the longterm prediction, our method still achieves the optimal results.