基于深度学习的入境物品自动化核酸采样系统

An Automated Nucleic Acid Sampling System for Entry Items Based on Deep Learning

本研究提出了一套基于深度学习的图像目标检测以及机械臂抓取技术的自动化核酸采样系统,采用“机械臂+摄像头”的方式对入境物品及其包装进行图像目标检测并进行拭子采样。本研究的目标检测框架以基于YOLOv5框架为蓝本,使用深度可分离卷积以及逐点卷积融合的方式作为新的主干网络,设计了更加轻量级的幻影卷积进行Neck结构中的特征提取,以此来改进检测速度。最后,通过比例-积分-微分(Proportion Integral Differential,PID)来控制机械臂进行对应物体的抓取或涂抹操作,以此保证机械臂在运行过程中的稳定性。在自建数据集上本研究修改后的模型检测指标达到了99.18%,在CPU上使用ONNX框架部署的速度达到了25 FPS,能够满足实时检测的需求。

In this study,an automated nucleic acid sampling system based on deep learning-based image object detection and robotic arm grasping technology was proposed,and the image object detection and swab sampling of entry items and their packaging were carried out by using the method of Robotic Arm plus Camera.Based on the YOLOv5 framework,the object detection framework in this study uses deep separable convolution and point-by-point convolution fusion as a new backbone network,and a more lightweight phantom convolution is designed for feature extraction in the Neck structure,so as to improve the detection speed.Finally,the Proportion Integral Differential(PID) is used to control the robotic arm to grasp or smear the corresponding object,so as to ensure the stability of the robotic arm during operation.On the self-built dataset,the modified model detection index reached 99.18%,and the deployment speed of the ONNX framework on the CPU reached 25 FPS,which can meet the requirements of real-time detection.