摘要
为防止非法人员进入变电站,恶意盗窃、破坏行为影响变电站正常运行,提出人员闯入智能检测系统。该系统以Hi3559A为计算单元,接入前端相机可见光数据,运行深度学习检测算法完成对监控区域的人员检测。针对精度优先的神经网络模型难以在资源受限的嵌入式平台上实现的问题,提出一种基于RetinaNet改进的轻量目标检测模型。该模型从网络骨架、特征融合方法和检测头3个部分进行优化,在减少网络冗余的同时,增加了检测模型的精度。此外,采用量化技术压缩模型,优化网络后处理步骤,使网络充分利用Hi3559A硬件资源,进一步减少前向推理耗时。实验结果表明,该系统的检测精度提升2.7%,达到87.0%,运行速率提升8.2倍,达到45.9帧/s,单帧运行时间21.8ms,满足设计要求。
To prevent illegal people from entering the substation whose theft and sabotage will affect the substation operation,we propose an intelligent human detection system.The control core of the system is Hi3559A,which collects images by controlling the front-end camera and detects human by deep learning.Considering the problem that the deep learning model is difficult to implement on the embedded platform,we propose a lightweight detection model based on RetinaNet optimized from the backbone,the feature fusion method,and the head that increases the accuracy while reducing the redundancy of the model.Besides,we compress the model by the quantization technology and further speed up by optimizing the post-processing that can efficiently utilize the neural network computing resources of Hi3559A.Finally,the precision is improved by 2.7%,reaching 87.0%;the speed is increased by 8.2 times to 45.9 fps;and the running time is 21.8ms that meets the design requirements.
作者
钱宇骋
朱太云
曹旭航
黎瑞
龚恩
王健
QIAN Yucheng;ZHU Taiyun;CAO Xuhang;LI Rui;GONG En;WANG Jian(State Grid Anhui Electric Power Research Institute,Hefei 230601,Anhui Province,China;State Key Laboratory of Multispectral Image Information Processing Technology(Huazhong University of Science and Technology),Wuhan 430070,Hubei Province,China)
出处
《电网技术》
EI
CSCD
北大核心
2021年第10期4181-4188,共8页
Power System Technology
基金
国家电网公司总部科技项目(5200-201927049A-0-0-00)。
