In recent years

human motion prediction is widely used in human-computer interaction

autonomous driving and other fields.A human motion prediction model named dense spatio-temporal separable graph convolution network (DSTS-GCN) based on spatio-temporal attention and dense graph convolution is proposed

which adopts an encoder-decoder structure.In the encoder part

a DSTS-GCN is constructed to improve the model’s motion feature extraction capability and solve the problem of inaccuracy long-term prediction.Secondly

a multi-head spatio-temporal attention mechanism is introduced to focus on the most important joints or body parts of the input action sequence

in order to improve the performance of the model.The decoder uses a temporal convolutional network (TCN) to decode and output the motion sequences of future frames.Experimental results on the Human3.6M dataset show that the DSTS-GCN model has strong robustness and stability.Compared with the baseline network

spatio-temporal separable graph convolution network (STS-GCN)

the average error of 80ms short-term prediction on the Human3.6M dataset is reduced by 10.8%

and the average error of 1000ms long-term prediction is reduced by 4.4%.The proposed model predicts future human motions more accurately.