【作者】 刘军；

【导师】 周文举；

【作者基本信息】 上海大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 智能相机由于其结构紧凑、集成度高和性能稳定等优点,在现代工业生产中具有广泛的应用前景。然而,目前市场上的智能相机大多为国外产品,价格昂贵,由于缺少模块化的专用算法和软件,智能相机的推广受到限制。本文研究来自于企业需求,研究探索了智能相机国产化设计方案,设计并实现了具有较高性价比的目标感知与识别一体化智能相机。本文基于智能相机组成原理,完成了智能相机的系统总体设计;提出了改进YOLOv5的轻量级目标检测算法,开发了智能相机的目标感知和目标识别模块;完成了智能相机硬件系统的搭建,实现了智能相机软硬件系统。基于本文研发的智能相机开发实现了三个不同场景下的典型应用案例。本文的主要研究工作概括如下:(1)提出了目标感知与识别一体化的智能相机总体设计方案。基于智能相机硬件组成原理,将智能相机系统划分为硬件系统和软件系统两部分,完成了智能相机系统总体方案设计。(2)设计了智能相机的目标感知模块和目标识别模块。根据视觉算法应用范围的不同,将目标感知模块划分为静态目标分割和运动目标动态分割两个子模块进行设计。根据视觉算法原理不同,将目标识别模块划分为基于传统图像处理的智能目标识别和基于深度学习的智能目标识别两个子模块进行设计。(3)提出了改进YOLOv5的轻量级目标检测算法SGCS-YOLOv5。在YOLOv5算法的基础上,对其主干网络和颈部网络进行了轻量化改进。在模型中加入了CA注意力机制,并使用了目标框优化的SIo U Loss来降低网络轻量化带来的检测精度损失。与YOLOv5n算法相比,本文提出的SGCS-YOLOv5模型在不降低m AP@0.5的情况下,模型的参数量降低了28.8%,浮点计算量降低了22.9%,实现了改进YOLOv5的轻量级目标检测算法SGCS-YOLOv5在智能相机平台的上的应用。(4)完成了智能相机系统的搭建。基于硬件开发平台,选取了合适的元器件,实现了智能相机的硬件系统。在硬件系统基础上,完成了操作系统和函数库的配置,实现了用户交互图形化控制界面。对智能相机的功能模块进行了测试,实现了智能相机系统软、硬件协调运行。(5)完成了智能相机在智能交通车辆检测、安瓿瓶外形歪头检测和堆叠钢板数量检测上的典型应用开发。在智能相机已有的视觉功能模块基础上,根据应用场景特点,开发相应检测功能。在智能交通车辆检测中,融合运动目标动态分割模块和SGCS-YOLOv5模块,实现了车流量的实时、高精度检测。在安瓿瓶外形歪头检测中,设计了基于静态目标分割模块的安瓿瓶区域提取方法和基于物体旋转中心提取的安瓿瓶外形歪头检测算法,实现了安瓿瓶外形歪头缺陷的简单、快速检测。在堆叠钢板数量检测中,基于静态目标分割模块和Prewitt特征,实现了堆叠钢板区域快速定位;针对堆叠钢板数量检测问题,提出了基于垂直投影法的检测算法,实现了准确度较高的堆叠钢板数量检测。经过实验与分析,本文设计的智能相机在三种典型应用场景下具有较好的检测效果。本文设计并实现了目标感知与识别一体化智能相机,所实现的智能相机具有体积小巧、功能丰富、操作方便、成本低等优点,可以应用于多种检测场景。更多还原

【Abstract】 Smart cameras,due to their compact structure,high integration,and stable performance,have a wide range of application prospects in modern industrial production.However,most smart cameras on the market are foreign products with high prices,and the promotion of smart cameras is limited due to the lack of modular specialized algorithms and software.This paper is based on the needs of enterprises and explores the localization design scheme of intelligent cameras,designing and implementing a cost-effective target perception and recognition integrated smart camera.Based on the composition principle of smart cameras,this paper completes the overall system design of the smart camera.An improved lightweight target detection algorithm based on YOLOv5,and develops the target perception and recognition modules of the smart camera.The hardware system of the smart camera is completed,and the hardware and software system of the smart camera is realized.Based on the smart camera developed in this paper,three typical application cases in different scenarios are implemented.The main research work of this paper is summarized as follows:(1)This paper presents an integrated design scheme for a smart camera system that combines target perception and recognition.The smart camera system is divided into hardware and software systems based on the principles of smart camera hardware composition,and the overall design of the smart camera system is completed.(2)The target perception module and target recognition module of the smart camera are designed.According to the different application scopes of visual algorithms,the target perception module is divided into static target segmentation and dynamic motion target segmentation sub-modules for design.Based on the different principles of visual algorithms,the target recognition module is divided into target recognition based on traditional image processing and intelligent target recognition based on deep learning for design.(3)An improved lightweight target detection algorithm,SGCS-YOLOv5,is proposed based on YOLOv5.The backbone network and neck network of YOLOv5 are improved for lightweight design.The Coordinate Attention is added to the model,and the SIo U Loss for target box optimization is used to reduce the detection accuracy loss caused by network lightweight design.Compared with YOLOv5 n,the proposed SGCSYOLOv5 model reduces the parameter size by 28.8% and the floating-point computation by 22.9% without significantly decreasing the m AP@0.5.Implemented the application of SGCS-YOLOv5,a lightweight target detection algorithm that improves YOLOv5,on the smart camera platform.(4)The construction of the smart camera system has been completed.Based on the hardware development platform,appropriate components were selected to implement the hardware system of the intelligent camera.On the foundation of the hardware system,the configuration of the operating system and function libraries was accomplished,realizing a user-friendly graphical control interface.The functional modules of the smart camera were tested,achieving the coordinated operation of the software and hardware systems of the intelligent camera system.(5)The smart camera is applied to typical applications such as intelligent traffic vehicle detection,ampoule bottle shape distortion detection,and stacked steel plate quantity detection.Based on the existing visual function modules of the smart camera,corresponding detection functions are developed according to the characteristics of the application scenarios.In intelligent traffic vehicle detection,a real-time and highprecision traffic flow detection method is achieved by integrating the dynamic motion target segmentation module and the SGCS-YOLOv5 module.In ampoule bottle shape distortion detection,a simple and fast detection method for ampoule bottle shape distortion defects is realized by designing a region extraction method based on the static target segmentation module and an algorithm based on the extraction of the object’s rotation center.In stacked steel plate quantity detection,a rapid location method for stacked steel plate areas is achieved based on the static target segmentation module and Prewitt features;for the detection of stacked steel plate quantity,a detection algorithm based on the vertical projection method is proposed,realizing a relatively accurate method for detecting the quantity of stacked steel plates.Through experiments and analysis,the designed intelligent camera demonstrates good detection performance in these three typical application scenarios.In this paper,an integrated smart camera for target perception and recognition is designed and implemented.The realized smart camera has the advantages of small size,rich functions,easy operation,and low cost,which can be applied to a variety of detection scenarios.更多还原