基于残差BP神经网络的机器人目标定位
Robot target positioning based on residual BP neural network
苏克 1闫人滏 1谢文宇 2崔倩雯 1张梦茜 1傅伯雄1
作者信息
- 1. 国网石家庄供电公司,河北石家庄 050051
- 2. 河北工业大学机械工程学院,天津 300401
- 折叠
摘要
为解决深度信息未知条件下的单目视觉机器人目标定位问题,文中基于相机的成像模型,提出一种基于三残差神经网络的三维坐标自解耦目标定位方法.首先,将YOLOv5s进行轻量化改进,实现目标的初步定位;然后,使用OpenCV提取目标的尺寸特征;接下来,根据相机的成像模型,将目标的尺寸特征与相机成像模型相融合,推导世界坐标系与像素坐标系和尺寸特征的函数关系;最后,采用残差网络避免梯度消失的优势,使用3个残差BP神经网络映射坐标转换函数,减小单个神经网络的工作量,获取目标的三维坐标信息.试验结果表明:该方法定位最大相对误差为0.747％.
Abstract
In this article,in order to ensure the target positioning of the monocular vision robot with unknown in-depth infor-mation,the three-dimensional coordinate self-decouple target-positioning method is proposed based on the three residual neural networks,with the help of the camera's imaging model.Firstly,YOLOv5s is subject to the lightweight improvement,and thus the target is positioned initially.Then,OpenCV is used to extract the target's size features.Next,they are fused with the camera's imaging model,so as to derive the functional relationship between the world coordinate system,the pixel coordinate system,and the size features.Finally,since the residual network is beneficial to avoid gradient disappearance,the three residual BP neural networks are used to map the coordinate-transformation function,which reduces the workload of a single neural network and ob-tains the information on the target's three-dimensional coordinate.The results show that the highest positioning error of this meth-od is 0.747％.
关键词
目标定位/图像处理/深度学习/BP神经网络Key words
target positioning/image processing/deep learning/BP neural network引用本文复制引用
基金项目
国家电网有限公司总部科技项目资助(kj2021-012)
出版年
2024
NETL
NSTL
万方数据