基于深度神经网络的人体运动识别系统设计

Design of human activity recognition system based on deep neural network

针对狭小空间下灾难救援人员运动状态感知困难的问题,提出一种基于卷积神经网络(CNN)+长短期记忆网络(LSTM)的人体运动识别(HAR)方法,设计了一种可运行于嵌入式微控制器单元(MCU)的HAR系统。系统使用佩戴于胸口的三轴加速度计及陀螺仪传感器作为数据输入,研究了卷积核数量和LSTM细胞数量对网络的影响,构建了HAR的深度学习模型。同时,针对存储占用、计算负荷和功耗对网络进行了优化与转译,并在嵌入式设备验证。结果表明:该系统可稳定运行于微控制器单元且对人体运动状态具有良好的识别精度。

Aiming at the problem of difficult state perception of disaster rescuers in confined spaces,a human activity recognition(HAR)method based on convolutional neural network(CNN)and long short-term memory(LSTM)network is proposed,and a HAR system which can run on embedded microcontroller unit(MCU)is designed.The system uses the three-axis accelerometer and gyroscope sensor worn on the chest as data input,the influence of the number of convolution kernels and LSTM cells on the network is studied,and the deep learning model for HAR is constructed.Meanwhile,the network is optimized and transcoded for storage occupancy,computational load and power consumption,and validated in embedded device.The results show that the system can run stably on MCU and has good recognition precision for human motion status.