【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。更多还原

【Abstract】 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.更多还原