基于增强型图卷积的骨架识别模型

Skeleton recognition model based on enhanced graph convolution

针对现有骨架动作识别主要采用双流框架,在提取时间空间以及通道特征方法上存在的问题,提出一个ADGCN,用于骨架动作识别。首先对骨架数据进行建模,分别将关节、骨骼及其关节和骨骼的运动信息输入到多流框架的单个流。然后将输入的数据传送到提出的有向图卷积网络中进行提取关节和骨骼之间的依赖关系,再利用提出的时空通道注意力网络(STCN),增强每层网络中关键关节的时间、空间以及通道的信息。最后将四个流的信息通过加权平均计算动作识别的精度,输出动作的预测结果。此模型在两个大型数据集NTU-RGB+D和Kinectics-Skeleton中进行训练和验证,验证的结果与基线方法DGNN(有向图神经网络)相比,在NTU-RGB+D数据集上,在两个交叉子集CS和CV上的准确率分别提升了2.43%和1.2%。在Kinectics-Skeleton数据集的top1和top5上的准确率分别提升了0.7%和0.9%。提出的ADGCN可以有效地增强骨架动作识别的性能,在两个大型数据集上的效果都有所提升。

In order to solve the problem of extracting temporal,spatial and channel features,this paper proposed an ADGCN for skeleton action recognition.At first,it modeled the skeleton data and fed the movement of joints,bones and joints and bones information into the multi-stream framework of a single stream.And then it transferred the input data to the directed graph convolution network to extract the dependencies between joints and bones,after that,used the space-temporal channel attention network(STCN)to strengthen the key joints in each layer of network temporal,space and channels of information.The last,it calculated the accuracy of action recognition by the weighted average of the information of four streams,output action prediction results.This model trained and verified in two large data sets NTU-RGB+D and Kinectics-Skeleton.Compared with baseline DGNN(directed graph neural network),the accuracy of this paper on two cross subsets CS and CV is improved by 2.43%and 1.2%,on NTU-RGB+D dataset.On Kinectics-skeleton dataset,the accuracy of top1 and top5 is increased by 0.7%and 0.9%,respectively.The proposed ADGCN can effectively enhance the performance of skeleton action recognition,and the effect is improved on both large data sets.