摘要
常规方法在识别可穿戴传感器获取的人体感知数据时,虽然能取得不错的分类性能,但是需要统计大量精确标记的数据,会消耗一定的人力和物力资源。针对上述问题,文中提出一种弱监督模式下的人体活动识别与定位算法。在卷积神经网络提取数据特征的过程中引入注意力机制,获取特征向量在时间维度的注意力权重,然后对权重序列进行中值滤波处理,找到局部最小值点作为区分活动与背景区域的分割阈值。在弱标签数据集上的实验结果表明,与基准模型CNN和DeepConvLstm相比,活动分类的准确率提升了3.64%和2.18%,当重叠度阈值取0.4时,活动定位的精确率达到了84.7%。
Although the conventional methods can achieve good classification performance when recognizing body sensing data obtained from wearable sensors,they require a large number of precisely⁃labeled data,Which may consume certain human and material resources.In view of the above problems,a weakly⁃supervised human activity recognition and localization algorithm is proposed in the paper.The attention mechanism is introduced in the process that the convolutional neural network is used to extract data features,and to obtain the attention weights of the feature vector in the time dimension.The weight sequence is processed with median filtering to find the local minimum point as the segmentation threshold to distinguish the action from the background region.The results of experiment on a weakly⁃labeled dataset show that,in comparison with the benchmark models of CNN and Deepconvlstm,the accuracy of activity classification of the proposed algorithm is increased by 3.64%and 2.18%,and the accuracy of activity localization can reach to 84.7%when the overlap threshold is 0.4.
作者
宋秀秀
周华
何军
胡昭华
SONG Xiuxiu;ZHOU Hua;HE Jun;HU Zhaohua(College of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China)
出处
《现代电子技术》
2021年第18期33-37,共5页
Modern Electronics Technique
基金
国家自然科学基金资助项目(61601230)
江苏省自然科学基金资助项目(BK20141004)。
关键词
人体活动识别
可穿戴传感器
数据处理
弱监督学习
注意力机制
中值滤波
human activity recognition
wearable sensor
data processing
weakly⁃supervised learning
attention mechanism
median filtering
