摘要
为了解决嵌入式手势识别系统的速度慢和功耗高的问题,提出基于深度学习处理单元(deep-learning processor unit,DPU)的手势识别系统设计方法。通过把DPU部署于现场可编程门阵列(field programmable gate array,FPGA)器件,再调用ResNet-50网络对手势图片进行识别,设计并实现了一套低功耗的手势识别系统。实验结果表明,系统在工作频率为150 MHz时识别准确率为97.7%,运行速率可以达到129 GOPS,其能效比为26.3 GOPS/W,优于一些现有嵌入式FPGA神经网络手势识别实现方法。
In order to solve the problems of low speed and high power consumption in the embedded gesture recognition system,a design scheme of gesture recognition system based on deep-learning processor unit(DPU)was proposed.By deploying the DPU on FPGA and applying ResNet-50 network to recognize gesture images,a low-power gesture recognition system was implemented.The experimental results show that the recognition accuracy is 97.7%and the running speed reaches 129 GOPS when the system operating frequency is 150 MHz,and the energy efficiency ratio is 26.3 GOPS/W,which is better than some existing embedded FPGA neural network gesture recognition methods.
作者
黎海涛
刘鸣
张帅
LI Haitao;LIU Ming;ZHANG Shuai(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Fan Gongxiu Honors College,Beijing University of Technology,Beijing 100124,China)
出处
《北京信息科技大学学报（自然科学版）》
2021年第3期1-7,共7页
Journal of Beijing Information Science and Technology University
基金
航空科学基金资助项目(2018ZC15003)。
关键词
深度学习处理单元
手势识别
神经网络
现场可编程门阵列
deep-learning processor unit
gesture recognition
neural network
field programmable gate array
