Neural networks based on graph convolution are simple and effective network structures that have better performance in some fields compared to traditional convolutional networks. However, when graph convolution networks (GCN) are applied to other fields, how to transform data or features into graph edges is a challenge needed to be studied. In previous studies of non-Euclidean spatial data such as relationship studies, people achieve graph construction by indicating the connection or not of nodes with binary values. However, when graph network is applied to some multi-sensor data scenarios, especially human motion tracking tasks, binarization is obviously not the best solution for the weights of edges. In addition, relying on manual graph construction may ignore some cross-space sensor relationships, which makes the update information of nodes far away slower or unavailable. Therefore, we try to use Pearson correlation coefficient (PCC) as the estimation of node correlation, and construct a complete graph to enhance the connectivity of nodes, which achieve faster signal transmission at remote nodes. After combining it with the temporal convolutional network (TCN), we propose a new network structure TGCN-P. Experimentally, this approach is more effective compared to other baseline models.