An Improved GRU Network for Human Motion Prediction
Weijie Yu, Rui Liu, Dongsheng Zhou, Qiang Zhang, Xiaopeng Wei · 2021
Human motion prediction is a research field with broad application prospects. With the development of deep learning, researchers have used advanced deep-learning algorithms in this field. This paper aims to combine GRU with 1D-CNN without increasing network parameters. In this paper, we use GRU to learn the continuity of human movement, and then use one-dimensional convolution networks to reduce the dimensions and to generate predicted actions. At last, we utilize the motion weight matrix which uses simple operations to get the weight from its own motion data, so as to improve the model's robustness. The test results on Human3.6M database show that our method gets a good result in short-term prediction and performs better than other GRU-based methods in long-term prediction. Other network using our ideas can also improve the prediction effect.