基于全局图遍历的ST-GCN人体行为识别算法

ST-GCN human action recognition algorithm based on global graph traversal

传统的时空图卷积网络(spatio-temporal graph convolutional network, ST-GCN)模型存在诸多缺陷,如空间图构造受预定义影响、忽略非相邻节点间信息的有效利用以及计算成本过高等.针对上述问题,基于ST-GCN模型提出一种采用节点全局图遍历的自适应人体行为识别方法.通过建立节点关联函数找出初始节点,再遍历拓扑状态库找出变化较大的前条链路筛选关键连接特征;建立当前节点与相邻和非相邻节点间的有效关联,在卷积过程中增加位运算操作用于舍弃关联度较小的关节点,以期在减少计算量的同时通过直接捕捉人体节点间的位置和距离信息调整运动关联度,增强算法的自适应性.实验结果表明,该算法较原始ST-GCN模型的识别精度高2%,且计算成本低,每秒浮点运算次数下降2.9×10^(9).

The traditional spatio-temporal graph convolutional network method has many defects,such as the influence of pre-definition on the construction of spatial graph,ignoring effective utilization of information between non-adjacent nodes,and high computational cost.Aiming at the present situation above,an adaptive human action recognition method of traversing the global graph of nodes is proposed based on the ST-GCN model.The initial node is found by establishing the node association function,the key connection features of the previous link are found by traversing the topology state library.The effective association between the current node and adjacent and non adjacent nodes is established.Meanwhile,bit operation is added to discard the joint points with small correlation degree in the convolution process,so as to adjust the motion correlation degree by directly capturing the position and distance information between human nodes while reducing the amount of computation,and enhance the adaptability of the algorithm.Experimental results show that compared with the original ST-GCN model,the recognition accuracy of the proposed algorithm is 2%higher,the calculation cost is lower,and the number of floating point operations per second is reduced by 2.9×10^(9).