文献类型：期刊文献

中文题名：基于空间特征的BI-LSTM人体行为识别

英文题名：Human Action Recognition Using BI-LSTM Network Based on Spatial Features

作者：付仔蓉[1];吴胜昔[1];吴潇颖[1];顾幸生[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237

年份：2021

卷号：47

期号：2

起止页码：225

中文期刊名：华东理工大学学报（自然科学版）

外文期刊名：Journal of East China University of Science and Technology

收录：CSTPCD;;Scopus;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家自然科学基金(61973120);上海汽车工业科技发展基金(1837)。

语种：中文

中文关键词：行为识别;骨骼数据;距离特征;角度特征;BI-LSTM

外文关键词：action recognition;skeletal data;distance feature;angle feature;BI-LSTM

摘要：随着微软Kinect等深度相机的出现,使用具有简洁性、鲁棒性和视图无关表示的3D骨架节点数据来识别人体行为的方法获得了很好的效果,但现有的针对骨骼序列数据的大多数学习方法缺少空间结构信息和详细的时空动态信息。利用双向长短期记忆网络(BI-LSTM)模型能长时间存储骨骼序列的特点获得丰富的双向时间信息对动作的顺序进行建模,同时从3D骨骼关节点坐标中提取关节点之间的相对距离特征和相对角度特征来加强空间结构特征,完成从骨骼数据中实现人体行为识别。该方法有效地进行了人体行为动作分类,提高了识别准确性。
With the advent of depth cameras such as microsoft Kinect,the method of recognizing human action via 3D skeleton node data with simplicity,robustness and view-independent representation has achieved quite good performance.However,most of the existing methods for skeleton sequence data lack spatial structure information and detailed temporal dynamics features.By means of the characteristics of BI-LSTM model with the long-term storage of skeleton sequences,rich bidirectional time information to model the sequence of actions is obtained.Meanwhile,the relative distance features and relative angle features between joint points are extracted from 3D bone joint point coordinates to strengthen the spatial structure features and realize the recognition of human action from the skeleton data.Finally,it is shown via the simulation results that this proposed method can effectively achieve the classification of human actions and improve the accuracy.