CWNU-RDA
收藏资源简介：
Human activity recognition (HAR) has attracted much attention. However, the existing HARs have shortcomings, such as few recognized activities, no identification, privacy leakage, and battery maintenance. Aiming at these shortcomings, this team has devised a body RFID skeleton that fully senses human activity and further proposes highly-accurate and fine-grained (total of 21 activities) HARs. The body RFID skeleton senses human activity by collecting tag response records of the skeleton node. The tag response records form the human activity dataset which is suitable for benchmarking the bound-RFID HAR. The proposed data standardization solves the problems caused by the differences in the size and the amount of the tag response signal feature data. Some body RFID skeleton-based HARs are proposed by utilizing machine learning and deep learning. Experiments show the validity of the body RFID skeleton-based HAR when faced with highly accurately recognizing fine-grained activities, that the body RFID skeleton-based HAR utilizing BiLSTM has the best accuracy (97.71%) among the proposed HARs, and that the body RFID skeleton-based HAR is easier to recognize activity which is expressed by RSSI and Phase. The proposed HARs can be used in consumer electronics such as healthcare. The dataset and source code are available at http://www.cwnuiot.net/BRS/.
人类活动识别（Human Activity Recognition，HAR）领域已受到学界广泛关注。然而现有HAR系统存在诸多不足：可识别活动类别有限、缺乏身份识别能力、存在隐私泄露风险，且电池维护存在难题。针对上述缺陷，本研究团队设计了一套可全面感知人体活动的人体射频识别（Radio Frequency Identification，RFID）骨骼传感系统，并进一步提出了高精度细粒度（涵盖21类活动）的HAR模型。该人体RFID骨骼传感系统通过采集骨骼节点的标签响应记录实现人体活动感知，此类标签响应记录构成了适用于绑定式RFID HAR基准测试的人体活动数据集。所提出的数据标准化方法解决了标签响应信号特征数据在维度与数量上的差异所引发的各类问题。研究团队结合机器学习与深度学习技术，构建了多款基于该人体RFID骨骼传感系统的HAR模型。实验结果表明，基于该系统的HAR模型在细粒度活动高精度识别任务中具备有效性；其中采用双向长短期记忆网络（Bidirectional Long Short-Term Memory，BiLSTM）的模型在本次提出的所有HAR模型中识别精度最高，达97.71%；且基于该系统的模型更易识别以接收信号强度指示（Received Signal Strength Indication，RSSI）和相位（Phase）表征的活动类别。本次提出的HAR模型可应用于医疗健康等消费电子领域。本数据集与源代码可通过http://www.cwnuiot.net/BRS/获取。
