基于可穿戴传感器的普适化人体活动识别

为了提高日常活动识别的准确性和自动化程度,减少人为干预,提出了利用可穿戴传感信号作为输入,通过深度神经网络进行人体活动识别的方法.首先,设计了普适环境下人体活动识别的系统架构,建立了一套加速度、生理信号等传感数据的采集系统;然后,对获取的传感数据进行降噪、加窗与归一化等预处理,并设计了长短时记忆递归神经网络来进行特征的自动提取和融合,从而实现活动识别.实验结果表明,该方法减少了对人工和专家知识的依赖,自动进行多模态传感器的融合,智能化程度高,分类效果好.

Human Daily Activity Recognition Based on Wearable Sensors

In order to improve accuracy and automation of human activity recognition, reduce manual involvement, a method is proposed to classify human activities by deep neural network based on heterogeneous wearable sensor signals. First, the architecture of human activity recognition system is designed, and a mobile application is developed to record sensor data such as acceleration, angular-rate and physiological signals in pervasive environment. Second, the sensor data is feed into well-designed long short-term memory network(LSTM) after preprocessing to recognize various activities by fusing inertial and physiological data. The test results show that this scheme can be independent of expert knowledge, and can classify complex human activities efficiently and accurately.