改进YOLO6D的目标姿态估计算法
Improved YOLO6D Target Attitude Estimation Algorithm
沈中华 1李涵 2程虎强 1甘增康3
作者信息
- 1. 桂林理工大学机械与控制工程学院,桂林 541000
- 2. 桂林理工大学机械与控制工程学院,桂林 541000;深圳职业技术大学智能制造技术研究院,深圳 518055
- 3. 深圳职业技术大学智能制造技术研究院,深圳 518055
- 折叠
摘要
针对三维空间下被遮挡和弱纹理目标物难以进行精确姿态估计的问题,提出了一种基于改进YOLO6D的目标姿态估计算法.首先,引入残差网络结构,解决了神经网络层数增加带来的梯度问题并加快模型收敛;其次,加入空间金字塔池化(SPP-CSP)模块使网络充分利用多尺度特征图信息来增强对目标物的特征提取.实验结果显示,改进后的网络在自建数据集上整体指标2D重投影上升了6.68%,5 cm5°上升了6.05%,在官方数据集Occlusion LineMOD上整体精度上升了 8.74%,有效提高了目标姿态估计的整体检测性能.
Abstract
This paper proposes an improved YOLO6D based target attitude estimation algorithm to address the problem of difficulty in accurately estimating the pose of occluded and weakly textured targets in three-dimensional space.Firstly,the residual network structure was introduced to solve the gradient problem caused by the increase in neural network layers and accelerate model convergence;Secondly,the addition of the spatial pyramid pooling(SPP-CSP)module enables the network to fully utilize multi-scale feature map information to enhance feature extraction of the target object.The experimental results show that the im-proved network has an overall increase of 6.68%in 2D reprojection and 6.05%in 5 cm5°on the self built dataset,and an overall accuracy increase of 8.74%on the official dataset Occlusion LineMOD,effectively improving the overall detection performance of target attitude estimation.
关键词
6D姿态估计/卷积神经网络/遮挡/特征提取Key words
6D attitude estimation/convolutional neural network/occlusion/feature extraction引用本文复制引用
基金项目
广东省教育厅重点领域专项(6022210111K)
校级科研启动基金(6022312029K)
出版年
2024
NETL
NSTL
万方数据