摘要
智能化环境和服务的重要基础在于能够对环境进行视觉建模,使其具有视觉识别和理解能力。为此,提出一种用于智能服务机器人的场景解析深度网络模型Shuffle-HRNet以实现自主移动和服务。设计一种Shuffle模块并引入HRNetV2网络,实现不同通道之间的信息交互,降低模型参数量并提高计算效率;提出一种聚合注意力感知模块,使网络关注每个通道中不同的有效特征信息、抑制不相关特征;在SmartLib数据集上对Shuffle-HRNet和主流分割方法进行了对比和消融实验。实验结果表明,Shuffle-HRNet能够对内部环境实现场景解析和准确分割。相比其他方法,Shuffle-HRNet具有更高的分割效率和更低的参数量,可部署于机器人以实现室内场景自主移动进而提供多元化服务。
The key foundation of intelligent service is to be able to visually model a environment and allow robots to possess visual recognition and parsing ability.Scene parsing can be widely applied in such fields as unmanned driving,image retrieval,and medical diagnosis.With the scene parsing technology,the semantic contours of targets in a scene can be detected and segmented.Then the specific semantics of the contours can be identified.Currently,ample research on intelligent libraries has been made based on new generation of information technologies including artificial intelligence.Intelligent robots in libraries can easily perform such tasks as identity recognition,reader guidance,book and informationretrieval,book inventory,reader information query,and intelligent consultation,which are of great value in the application research of intelligent libraries.How to use visual systems to achieve scene parsing and then navigate and act autonomously to achieve intelligent services has important research significance.However,intelligent warehousing for intelligent libraries,automatic inventory robots,and navigation robots,etc.still largely rely on infrared rays,ultrasound,Wi-Fi,Bluetooth,and other technologies for modeling.True intelligence is still far away.In addition,the varied indoor layout of intelligent libraries,the more complex environment and high reader mobility pose other challenges.Existing visual scene parsing technologies are still confronted with issues in terms of high resolution,low latency,lightweight,and edge computing deployment.In recent years,the attention mechanism has developed rapidly in the field of computer vision based on deep learning.By imitating the human visual and cognitive systems,it enables deep learning models to selectively focus on relevant data,thereby efficiently allocating limited computational resources and improving efficiency.This paper presents a scene parsing method Shuffle-HRNet of intelligent library based on convergent attention perception,which allows intelligent service robo
作者
张岩
孙英伟
ZHANG Yan;SUN Yingwei(Library,Qingdao University of Science and Technology,Qingdao 266000,China;College of Mechanical and Electrical Engineering,Qingdao University of Science and Technology,Qingdao 266000,China)
出处
《重庆理工大学学报（自然科学）》
北大核心
2023年第10期136-145,共10页
Journal of Chongqing University of Technology：Natural Science
基金
山东省自然科学基金项目(ZR2019MEE066)。
关键词
智慧图书馆
场景解析
聚合注意力感知
计算机视觉
人工智能
smart library
scene parsing
convergent attention perception
computer vision
artificial intelligence
