户外智能随从机器人系统设计  

Design of Intelligent Follower Robot Used Outdoor

作　　者：邓开连[1] 朱华章 燕帅[1] DENG Kailian;ZHU Huazhang;YAN Shuai(College of Information Science and Technology,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620

出　　处：《实验室研究与探索》2023年第2期98-102,共5页Research and Exploration In Laboratory

基　　金：2021年度国家级大学生创新创业训练计划项目(G180910418);2021年中国纺织工业联合会高等教育教学改革研究项目(2021BKJGLX140);上海市高等教育学会2021年度规划研究项目(Z2-10)。

摘　　要：设计了一款基于计算机视觉的随从机器人。与常规智能小车相比,该智能机器人实现了多个神经网络协同工作,能够跟随主人在户外移动,并且按照主人做出的指令执行相应的动作,具备人脸识别、基于行为识别的摔倒检测等功能。该智能机器人使用FaceNet进行面部特征提取;采用YOLOV4+DeepSort组合实现用户跟踪;基于BlazePose设计了手势命令;设计并训练了基于LSTM的行为识别网络。实验结果表明,在对人体行走、坐下、摔倒3类动作分类的准确率分别达到了93.6%、96.7%和97.8%,能够有效地检测摔倒姿势。系统在达到预测的准确性同时占用更少的运算资源,将模型部署在搭载了拥有512个CUDA核心的Volta^(TM)GPU的JetsonNX上,运行帧率达到了15帧/s以上,具有良好的实时性。This paper designs a follower robot based on computer vision.Compared with the conventional intelligent car,this intelligent robot realizes the cooperative work of multiple neural networks.It can follow the owner to move outdoor and perform corresponding actions according to the instructions delivered by the owner.It is equipped with the functions of face recognition,falling detection based on behavior recognition,etc.FaceNet is used for facial feature extraction;YOLOV4+DeepSort combination is used to achieve user tracking.Gesture commands are designed based on BlazePose;and a LSTM-based behavior recognition networkis designed and trained.The experimental results show that in the classification of human walking,sitting and falling,the accuracy rates reached 936%,967%and 978%,respectively.It can effectively detect falling postures.The system can achieve a high accuracy of prediction while occupying less computing resources,the frame rate of running on JetsonNX with Volta^(TM)GPU which has 512 CUDA core inside reaches more than 15 frames per second,which has good real-time performance.

关 键 词：户外跟随 智能机器人 系统设计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]