基于BCDV的视觉传感器算法  

Design of Internet of Things Vision Sensor Based on Attitude Recognition

作　　者：黄富嗣 陈湘骥[1] HUANG Fu-si;CHEN Xiang-ji(School of Mathematics and Information,South Agricultural University,Guangzhou 510642)

机构地区：[1]华南农业大学数学与信息学院,广州510642

出　　处：《现代计算机》2020年第12期93-96,共4页Modern Computer

基　　金：广东省大学生创新创业训练计划项目(No.201810564102)。

摘　　要：随着物联网的普及,视觉传感器的应用越来越广,部署的数量越来越多,对网络带宽和服务器的处理能力带来巨大压力,为了减轻服务器负载,在视觉传感器上进行边缘计算成为研究的热点。深度学习算法可以很好地对图像进行更深层次的处理、压缩信息、减轻带宽负载,但在实际应用中发现,一般嵌入式设备运行深度学习算法达不到实时处理的帧数。该算法通过融合背景差分和三帧差算法,通过计算分块质心位移方差输出多人运动趋势信息、姿态信息,在树莓派上实现实时图像处理,平均帧数达到21帧。With the popularization of the Internet of Things,the application of vision sensors is becoming more and more widespread,and the number of deployments is increasing,which brings great pressure on network bandwidth and server processing capabilities.In order to reduce serv⁃er load,edge computing is performed on vision sensors Become a research hotspot.Deep learning algorithms can perform deeper processing on images,compress information,and reduce bandwidth loads.However,in actual applications,it has been found that general embedded devices cannot run deep learning algorithms to achieve the number of frames processed in real time.In this paper,the background differ⁃ence and three frame difference algorithms are combined.By calculating the block centroid displacement variance and outputting multiperson movement trend information and attitude information,real-time image processing is implemented on the Raspberry Pi,and the aver⁃age frame number reaches 21 frames.

关 键 词：嵌入式 视觉传传感器 姿态识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP212[自动化与计算机技术—计算机科学与技术]