基于边缘设备轻量化行为识别算法  

Lightweight behavior recognition algorithm based on edge devices

作　　者：郑永生 肖军 温高能 雷磊[4] 彭勃兴 文润玉 ZHENG Yongsheng;XIAO Jun;WEN Gaoneng;LEI Lei;PENG Boxing;WEN Runyu(China National Aviation Fuel Group Haixin Shipping Co.,Ltd.,Shanghai 200051,China;North China Company of China National Aviation Fuel Supply Co.,Ltd.,Beijing 100102,China;Aerospace Shenzhou Intelligent System Technology Co.,Ltd.,Beijing 100029,China;College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]中国航油集团海鑫航运有限公司,上海200051 [2]中国航空油料有限责任公司华北公司,北京100102 [3]航天神舟智慧系统技术有限公司,北京100029 [4]四川大学电子信息学院,四川成都610065

出　　处：《现代电子技术》2023年第23期137-143,共7页Modern Electronics Technique

摘　　要：针对机场加油车等某些生产场景下工作人员的行为得不到实时性监督的问题,提出一种可部署至边缘设备轻量化加油员行为识别算法。该算法首先使用基于YOLOv5s改进的目标检测网络进行快速人体检测;再使用IoU和直方图相似度相结合的跟踪算法对检测到的人体目标进行跟踪,由跟踪得到的序列图像通过轻量级的姿态估计网络预测出人体的骨骼关键点序列数据;最后将骨骼关键点序列数据输入到6层的全连接网络分类器中进行动作分类,判断加油员动作是否规范完成。实验数据表明:该算法大大减少了网络权重和计算量,其中改进后的人体检测网络YOLOv5-mini在边缘设备比特大陆Sophon SE5上单帧检测速度可达18 ms;在实际场景数据集上,算法行为检测准确率可达95.92%。In order to solve the problem that the behavior of the staff in some production scenarios,such as airport refueling trucks,can not be monitored in real time,a lightweight refueling staff behavior recognition algorithm that can be deployed to edge devices is proposed.In the algorithm,the improved object detection network based on YOLOv5s is used to perform fast human detection first,and then the tracking algorithm combining IoU and histogram similarity is used to track the detected human objects,so as to predict the sequence data of human skeleton key points by the lightweight pose estimation network based on the tracked sequence images,and the sequence data of skeleton key points is input into the 6⁃layer fully⁃connected network classifier for action classification,so as to judge whether the fuel dispenser's actions are completed in a standard manner.The experimental data show that the algorithm can greatly reduce the network weight and computation.The improved human detection network YOLOv5⁃mini can achieve a single frame detection speed of 18 ms on the edge device BITMAIN Sophon SE5,and the accuracy of behavior detection of this algorithm can reach 95.92%on the actual scene dataset.

关 键 词：目标检测 YOLOv5 目标跟踪 骨骼关键点 行为识别 轻量化 

分 类 号：TN911-34[电子电信—通信与信息系统] TP311[电子电信—信息与通信工程]