摘要
为更好地对动作进行分类,提出基于推理网络的人体动作识别算法。通过Faster RCNN提取以人为主要区域、以场景信息为附加区域的特征信息,将其输入到LSTM中进行边框回归以及动作分类,通过结合Faster RCNN和LSTM获得动作的空间特征和时间特征,得到更精确的动作分类。在公认的两个数据集上进行实验,UCF-101数据集上精确度达到了93.6%,HMDB-51数据集中精确度达到了66.2%。
To classify the human action better,a human action recognition algorithm based on inference network was proposed.The feature information of the human main area and the scene information as the additional area was extracted using Faster RCNN.And it was inputted into LSTM for frame regression and action classification.By combining Faster RCNN and LSTM,rich spatial features and temporal features in the video were obtained to get more accurate motion classification.Experiments were performed on two recognized data sets,with an accuracy of 93.6%on the UCF-101 data set and an accuracy of 66.2%on the HMDB-51 data set.
作者
葛鹏花
智敏
GE Peng-hua;ZHI Min(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,China)
出处
《计算机工程与设计》
北大核心
2021年第3期853-858,共6页
Computer Engineering and Design
基金
内蒙古自然科学基金项目(2018MS06008)。
关键词
推理网络
人体动作识别
上下文信息
快速区域卷积神经网络
长短时记忆网络
inference network
human action recognition
contextual information
faster regional convolutional neural network(Faster RCNN)
long short term memory(LSTM)
