Aiming at the problem of how to build a high-performance model with low number of parameters and calculation, this paper proposed a lightweight multi-information graph convolutional neural network(LMI-GCN). LMI-GCN performed information fusion by encoding four different information of joint coordinates, joint speed, bone edge, and bone edge speed into a high-dimensional space, and introduced a multi-channel adaptive graph that could aggregate important features and a shunt temporal convolution block to reduce the amount of model parameters. At the same time, this paper proposed a random pool data preprocessing method. Compared with the baseline method SGN(semantics-guided neural networks) on the NTU-RGB+D120 dataset, the two evaluation settings cross-subject and cross-setup increase by 5.4% and 4.7%. Experimental results show that the performance of the LMI-GCN is better than that of the SGN.

[2021-11-22] Accepted Paper /doc/2021.07.0329/v1

[2022-04-05] Printed Article /doc/2021.07.0329/v2