基于路径签名的改进时空图卷积网络

针对时空图卷积网络ST-GCN中GCN的关节邻接图不易学习远端关节之间的语义信息和TCN在描述时间信息方面存在不足的问题，引入了数字签名预处理来增强数据，提出了基于路径签名的改进时空图卷积网络SSIT-GCN。首先将关节位置坐标的时间序列输入签名层进行数据预处理，在该层时间序列通过嵌入算法被转换为多维路径，将其划分为多条路径并计算每条路径的签名特征；其次重新设计GCN的关节邻接矩阵，并用反卷积来代替补零，以保持TCN的尺寸不变，还引入1×1的卷积核增加非线性来改进ST-GCN,得到改进时空图卷积网络SIT-GCN;最后用签名特征代替原始数据输入SIT-GCN，得到最终的输出结果。实验结果表明，基于路径签名的改进时空图卷积网络大大提高了训练精度，缩短了训练时间，对动态手势识别有较好的识别能力和识别速度。

Signature spatial improved temporalgraph convolutional network

Aiming at the problem that the joint adjacency graph of GCN in Spatial Temporal Graph Convolutional Network (ST-GCN) is not easy to learn the semantic information between distal joints and that TCN is insufficient in describing time information, the digital signature preprocessing is introduced to enhance data, and a signature spatial improved temporal graph convolutional network (SSIT-GCN) is proposed. Firstly, the time series of human joint locations are input into the signature layer to preprocess the data, and they are transformed into a multi-dimensional path by various embedding algorithms. The multi-dimensional path is divided into multiple paths and the signature features of each path are calculated. Secondly, the adjacency matrix of GCN is redesigned, and deconvolution is used to replace zero padding to keep the size of TCN unchanged, and a 1×1 convolution kernel is also introduced to increase the nonlinearity to improve ST-GCN, so as to obtain spatial improved temporal graph convolutional network(SIT-GCN). Finally, the original data is replaced by signature features that is input into SIT-GCN to obtain the final result. The experimental results show that the signature-based SSIT-GCN greatly improves the training accuracy, reduces the training time, and has better recognition ability and speed for dynamic gesture recognition.