基于FPGA的人体行为识别系统的设计

为实现边缘端人体行为识别需满足低功耗、低延时的目标,本文设计了一种以卷积神经网络(CNN)为基础、基于可穿戴传感器的快速识别系统.首先通过传感器采集数据,制作人体行为识别数据集,在PC端预训练基于CNN的行为识别模型,在测试集达到93.61%的准确率.然后,通过数据定点化、卷积核复用、并行处理数据和流水线等方法实现硬件加速.最后在FPGA上部署识别模型,并将采集到的传感器数据输入到系统中,实现边缘端的人体行为识别.整个系统基于Ultra96-V2进行软硬件联合开发,实验结果表明,输入时钟为200 M的情况下,系统在FPGA上运行准确率达到91.80%的同时,识别速度高于CPU,功耗仅为CPU的1/10,能耗比相对于GPU提升了91%,达到了低功耗、低延时的设计要求.

Design of human activity recognition system based on FPGA

In order to achieve the goal of low power consumption and low latency for edge-end human activity recognition,this paper designs a fast recognition system based on wearable sensors and Convolutional Neural Networks (CNNs).First,the system collects data through sensors to make a human activity recognition dataset,and pre-trains a CNN-based behavior recognition model on the PC side,which achieves an accuracy of 93.61% on the test set.Then,hardware acceleration is realized through methods such as data fixed point,convolution kernel multiplexing,parallel processing of data,and pipeline.Finally,the recognition model is deployed on the FPGA,and the collected sensor data are input into the system to realize the recognition of human activity at the edge.The whole system is developed jointly with hardware and software based on Ultra96-V2.The experimental results show that when the input clock is 200 M,the system runs on FPGA with an accuracy of 91.80%;the proposed system is superior to CPU in recognition speed as well as power consumption,specifically,the power consumption is only one-tenth of CPU consumed,and energy consumption ratio is 91% higher than that of GPU.It can be concluded that the FPGA-based human activity recognition system meets the design requirements of low power consumption and low delay.