基于图卷积神经网络的人体行为识别方法研究

The Study of Graph Convolutional Network for Human Action Recognition

【目的】随着互联网、人工智能、云计算等技术的快速发展,视频、图像等文件的数据量也日益增长,探究选择合适的关键技术对这些数据中蕴含的有用信息进行挖掘。【方法】基于图卷积神经网络对人体3D骨架关键节点各种行为进行识别和研究,在MSR-Action 3D数据集中对骨架特征增强图卷积网络模型与近5年常用的人体行为识别算法识别行为的准确率进行对比。【结果】研究结果表明,骨架特征增强图卷积网络模型对人体行为识别的准确率最高。【结论】骨架特征增强图卷积网络模型为后期智慧交通领域的人体异常行为识别提供有价值的参考。

[Purposes]With the rapid development of Internet,artificial intelligence,cloud computing and other technologies,the amount of data such as videos and images is also increasing.[Methods]Based on the graph convolutional neural network,various behaviors of key nodes of human 3D skeleton are identified and studied.In the MSR-Action 3D data set,the accuracy of the skeleton feature enhanced graph convolutional network model and the human behavior recognition algorithm commonly used in the past five years is compared.[Findings]The results show that the skeleton feature enhanced graph convolutional network model has the highest accuracy in human behavior recognition.[Conclusions]The skeleton feature enhanced graph convolutional network model provides a valuable reference for human abnormal behavior recognition in the field of intelligent transportation.