Abstract
In the task of human motion recognition, the overall action span is changeable, and there may be an inclusion relationship between action semantics. This paper proposes a novel multi-scale time sampling module and a deep spatiotemporal feature extraction module, which strengthens the receptive field of the feature map and strengthens the extraction of spatiotemporal-related feature information via the network. We study and compare the performance of three existing multi-channel fusion methods to improve the recognition accuracy of the network on the open skeleton recognition dataset. In this paper, several groups of comparative experiments are carried out on two public datasets. The experimental results show that compared with the classical 2s-AGCN algorithm, the accuracy of the algorithm proposed in this paper shows an improvement of 1% on the Kinetics dataset and 0.4% and 1% on the two evaluating indicators of the NTU-RGB+D dataset, respectively.
Funder
key special project of the National Key R&D Program
Subject
Fluid Flow and Transfer Processes,Computer Science Applications,Process Chemistry and Technology,General Engineering,Instrumentation,General Materials Science
Reference42 articles.
1. Unsupervised learning of depth estimation based on attention model and global pose optimization
2. Visual perception of biological motion and a model for its analysis
3. Semi-supervised classification with graph convolutional networks;Kipf;arXiv,2016
4. Spectral networks and locally connected networks on graphs;Bruna;arXiv,2013
Cited by
31 articles.
订阅此论文施引文献
订阅此论文施引文献，注册后可以免费订阅5篇论文的施引文献，订阅后可以查看论文全部施引文献