摘要
为了解决在计算与存储资源受限的嵌入式设备中实现人脸追踪与识别的问题,设计了一种基于Maixduino AI K210开发板的YOLOX-Nano与MobileFaceNet轻量级人脸追踪和识别系统方案。将输入图像的长宽缩减至原来的7/13,使YOLOX-Nano运算量从1.08 GFLOPS缩减到0.31 GFLOPS,将MobileFaceNet网络的激活函数替换为Maixduino AI K210开发板KPU支持的LeakyRelu算子。通过将图像输入缩减后的YOLOX-Nano得到人脸的定位,并将定位信息转化成角度发送给舵机,舵机转动使人脸始终保持在图像中心位置。通过改进后的MobileFaceNet网络进行人脸识别,实现人脸追踪与识别。通过设计多种实际情形进行实验,验证了该系统在实际使用时具有较好的稳定性和准确性。
In order to solve the problem of the realization of face tracking and recognition in embedded devices with limited computing and storage resources,a YOLOX⁃Nano and MobileFaceNet based lightweight human face tracking and recognition system design strategy is implemented in Maixduino AI K210 evaluation board.The length and width of the input image are compressed into 7/13 of the original.The YOLOX⁃Nano computation can be reduced from 1.08 GFLOPS to 0.31 GFLOPS.The activation function of the MobileFaceNet network is replaced with the LeakyRelu operator which is supported by the KPU of Maixduino AI K210 evaluation board.The image is input into the reduced YOLOX⁃Nano to get the positioning of the face.The positioning information is converted into an angle which is sent to a steering gear.The steering gear rotates to keep the face in the center of the image.Human face recognition can be conducted via the improved MobileFaceNet.Human face tracking and recognition can be achieved in this process.This paper validates the stability and accuracy of the system in practical use through a variety of experiments in practical situations.
作者
车佳祺
许晓荣
梁颢铭
CHE Jiaqi;XU Xiaorong;LIANG Haoming(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《电子设计工程》
2022年第14期58-63,共6页
Electronic Design Engineering
基金
国家留学基金委公派访问学者项目(202108330152)
杭州电子科技大学“优秀骨干教师支持计划”人才项目。
