文献类型：期刊文献

中文题名：基于CNN与双向LSTM的行为识别算法

英文题名：Action recognition algorithm based on CNN and bidirectional LSTM

作者：吴潇颖[1];李锐[1];吴胜昔[1]

机构：[1]华东理工大学化工过程先进控制和优化技术教育部重点实验室

年份：2020

卷号：41

期号：2

起止页码：361

中文期刊名：计算机工程与设计

外文期刊名：Computer Engineering and Design

收录：CSTPCD;;北大核心:【北大核心2017】；

基金：国家自然科学基金项目（61573144）;上海汽车工业科技发展基金项目（1837）

语种：中文

中文关键词：行为识别;体感摄像机;骨骼;卷积神经网络;双向长短期记忆网络

外文关键词：action recognition;Kinect;skeleton;CNN;Bi-LSTM

摘要：针对传统行为识别依赖手工提取特征,智能化程度不高,识别精度低的问题,提出一种基于3D骨骼数据的卷积神经网络(CNN)与双向长短期记忆网络(Bi-LSTM)的混合模型。使用3D骨骼数据作为网络输入,CNN提取每个时间步的3D输入数据间的空间特征,Bi-LSTM更深层地提取3D数据序列的时间特征。该混合模型自动提取特征完成分类,实现骨骼数据到识别结果的端对端学习。在UTKinect-Action3D标准数据集上,模型的识别率达到97.5%,在自制Kinect数据集上的准确率达到98.6%,实验结果表明,该网络有效提高了分类准确率,具备可用性和有效性。
Aiming at the problems that traditional action recognition relies on manual feature extraction,which is not intelligent enough and has low recognition accuracy,a hybrid model of convolutional neural network(CNN)and bidirectional long-term and short-term memory network(Bi-LSTM)based on 3D skeleton data was proposed.3D skeleton data were used as the network input.CNN extracted the spatial features between the 3D input data for each time step,and Bi-LSTM extracted the temporal features of 3D data series more deeply.The hybrid model automatically extracted features to complete classification and achieved end-to-end learning from skeleton data to recognition results.On the UTKinect-Action3D standard dataset,the recognition rate of the model is 97.5%,and the accuracy rate of the self-made Kinect dataset is 98.6%.Experimental results show that the network effectively improves the classification accuracy,which has availability and effectiveness.