一种基于QueryDet-YOLO的小目标实时检测定位系统  

A small object detection and location system based on QueryDet-YOLO

作　　者：王红星[1] 刘青[1] 吴欢娣 梁伟 邱舒鑫 吕凌波 刘超 张夕军 WANG Hongxing;LIU Qing;WU Huandi;LIANG Wei;QIU Shuxin;LV Lingbo;LIU Chao;ZHANG Xijun(School of Mechanical Engineering,Nanchang Institute of Technology,Nanchang 330099,China;Library,Nanchang Institute of Technology,Nanchang 330099,China;Yaohu College,Nanchang Institute of Technology,Nanchang 330099,China;Haier Robot Technology(Qingdao)Co.,Ltd.,Qingdao,266032,China)

机构地区：[1]南昌工程学院机械工程学院,江西南昌330099 [2]南昌工程学院图书馆,江西南昌330099 [3]南昌工程学院瑶湖学院,江西南昌330099 [4]海尔机器人科技(青岛)有限公司,山东青岛266032

出　　处：《南昌工程学院学报》2023年第6期71-75,共5页Journal of Nanchang Institute of Technology

基　　金：江西省教育厅科技项目项目(GJJ190941;GJJ201917);安徽省教育厅自然科学基金项目(KJ22020A0364);江西省教育厅自然科学基金项目(20212BAB204039);瑶湖学院学生创新基金项目“面向协同跟踪任务共融类人机器人运动优化研究”。

摘　　要：提高小目标检测最直接的方法是采用更高分辨率图像,其缺点是计算成本随着图像像素的增加而急剧增加,并产生冗余计算。本文基于QueryDet-YOLO模型,采用Kinect传感器设计了一种小目标实时检测定位系统。在ROS(Robot Operating System)系统下,Kinect传感器利用OpenCV,通过查询机制和映射关系预测小目标物体在低分辨率特征上的位置,然后利用预测位置在HD(High Definition)下进行稀疏卷积,精确计算出小目标检测结果,并配合实时点云数据输出小目标三维坐标信息,实现小目标实时精确定位。实验结果表明,与直接在高分辨率下检测相比,本文设计的小目标实时检测定位系统减少了计算量、加快了检测速度和提高了小目标实时检测效率。The most direct way to improve small target detection is to use higher resolution images.The disadvantage is that the computational cost increases sharply with the increase of image pixels,and higher redundant calculations are generated.This article is based on the QueryDet Yolo model and uses Kinect sensors to design a real-time detection and positioning system for small targets.In the ROS(Robot Operating System),Kinect sensors use OpenCV to predict the position of small target objects on low resolution features through query mechanisms and mapping relationships.Then,sparse convolution is performed using the predicted position in HD(High Definition)to accurately calculate the small target detection results,and real-time point cloud data is used to output the three-dimensional coordinate information of the small target,achieving real-time and accurate positioning of the small target.The experimental results show that compared with direct detection at high resolution,the small target real-time detection and positioning system designed in this paper reduces computational complexity,accelerates detection speed,and improves the efficiency of small target real-time detection.

关 键 词：小目标检测 YOLO QueryDet-YOLO KINECT 点云定位 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]