基于ST-GCN警用巡逻机器人警情识别系统设计  

Design of police situation recognition system for police patrol robots based on ST-GCN

作　　者：胡丽军 吴燕玲 宋全军 徐湛楠 HU Lijun;WU Yanling;SONG Quanjun;XU Zhannan(College of Electrical Engineering and Automation,Anhui University,Hefei 230601,China;Institute of Intelligent Machinery,Hefei Institute of Physical Sciences,Chinese Academy of Sciences,Hefei 230031,China)

机构地区：[1]安徽大学电气工程与自动化学院,安徽合肥230601 [2]中国科学院合肥物质科学研究院智能机械研究所,安徽合肥230031

出　　处：《传感器与微系统》2023年第6期78-81,共4页Transducer and Microsystem Technologies

基　　金：安徽省重点研究和开发计划资助项目(201904d07020007,202004a07020051);中国科学院合肥物质科学研究院院长基金资助项目(YZJJ2020QN17)。

摘　　要：针对现有警用巡逻机器人警情识别系统识别种类单一、识别率较低等问题,基于时空—图卷积网络(ST-GCN)和OpenPose算法的融合,面向跌倒(fall)、打砸(smash)和推搡(push)3种警情,设计了警用巡逻机器人警情识别系统。通过在真实场景测试,3种警情识别率分别为:跌倒85%,打砸80%,推搡83%;实时识别帧率为10 fps。实验结果表明:所设计的警情识别系统可以实现多种警情的实时、准确、可靠识别,具有较高的应用价值。Aiming at the problems of single recognition species and low recognition rate of police situation recognition systems of existing police patrol robots,based on fusion of spatio-temporal graph convolutional network(ST-GCN)and OpenPose algorithms,and oriented to three police situations:Fall,Smash and Push,design a police situation recognition system for police patrol robots.Through testing in real scenes,the three typical police situation recognition rates are:Fall:85%,Smash:80%,Push:83%,respectively,and the real-time recognition frame rate is 10 fps.The experimental results show that the designed police situation recognition system can realize the real-time,accurate and reliable identification of a variety of police situations,and has high application value.

关 键 词：时空—图卷积网络 OpenPose算法 警情识别 警用巡逻机器人 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置] TP391.41[自动化与计算机技术—控制科学与工程]