针对在复杂场景下对弱纹理目标位姿估计的准确性和实时性问题,提出基于筛选学习网络的六自由度(6D)目标位姿估计算法。首先,将标准卷积替换为蓝图可分离卷积(BSConv)以减少模型参数,并使用GeLU(Gaussian error Linear Unit)激活函数,能...针对在复杂场景下对弱纹理目标位姿估计的准确性和实时性问题,提出基于筛选学习网络的六自由度(6D)目标位姿估计算法。首先,将标准卷积替换为蓝图可分离卷积(BSConv)以减少模型参数,并使用GeLU(Gaussian error Linear Unit)激活函数,能够更好地逼近正态分布,以提高网络模型的性能;其次,提出上采样筛选编码信息模块(UFAEM),弥补了上采样关键信息丢失的缺陷;最后,提出一种全局注意力机制(GAM),增加上下文信息,更有效地提取输入特征图的信息。在公开数据集LineMOD、YCB-Video和Occlusion LineMOD上测试,实验结果表明,所提算法在网络参数大幅度减少的同时提升了精度。所提算法网络参数量减少近3/4,采用ADD(-S) metric指标,在lineMOD数据集下较Dual-Stream算法精度提升约1.2个百分点,在YCB-Video数据集下较DenseFusion算法精度提升约5.2个百分点,在Occlusion LineMOD数据集下较像素投票网络(PVNet)算法精度提升约6.6个百分点。通过实验结果可知,所提算法对弱纹理目标位姿估计具有较好的效果,对遮挡物体位姿估计具有一定的鲁棒性。展开更多
The application of high-performance imaging sensors in space-based space surveillance systems makes it possible to recognize space objects and estimate their poses using vision-based methods. In this paper, we propose...The application of high-performance imaging sensors in space-based space surveillance systems makes it possible to recognize space objects and estimate their poses using vision-based methods. In this paper, we proposed a kernel regression-based method for joint multi-view space object recognition and pose estimation. We built a new simulated satellite image dataset named BUAA-SID 1.5 to test our method using different image representations. We evaluated our method for recognition-only tasks, pose estimation-only tasks, and joint recognition and pose estimation tasks. Experimental results show that our method outperforms the state-of-the-arts in space object recognition, and can recognize space objects and estimate their poses effectively and robustly against noise and lighting conditions.展开更多
Conventionally, image object recognition and pose estimation are two independent components in machine vision. This paper presented a simple but effective method KNNSNG, which tightly couples these two com ponents wit...Conventionally, image object recognition and pose estimation are two independent components in machine vision. This paper presented a simple but effective method KNNSNG, which tightly couples these two com ponents within a single algorithm framework. The basic idea of this method came from the bionic pattern recog nition and the manifold ways of perception. Firstly, the shortest neighborhood graphs （SNG） are established for each registered object. SNG can be regarded as a covering and triangulation for a hypersurface on which the training data are distributed. Then for recognition task, the deter mined test image lies on which SNG by employing the parameter ＂k＂, which can be calculated adaptively. Finally, the local linear approximation method is adopted to build a local map between highdimensional image space and lowdimensional manifold for pose estimation. The projective coordinates on manifold can depict the pose of object. Experiment results manifested the effectiveness of the method.展开更多
文摘针对单帧RGB-D图像进行物体六自由度位姿估计时,在物体遮挡、光线情况不良、低纹理情况下性能不佳的问题,本文设计了一种基于多网络特征融合(颜色特征提取网络和点云特征提取网络)的深度学习网络.首先,使用颜色特征提取网络提取RGB图像中的纹理特征,使用点云特征提取网络计算深度图中的点云特征,进行几何特征与纹理特征计算后,回归计算点云的关键点投票及实例语义信息.然后,通过投票聚类方式计算每个实例的所属类别和关键点位置.将RGB-D图像中的颜色信息与几何信息分别计算,由于后续操作需要充分考虑像素及点云的局部信息与全局信息,分别使用改进后的残差神经网络和RIPoint(residuals inverted point)网络提取数据特征.采用神经网络中的特征融合方法将颜色信息与几何信息充分提取,为后续模块提供更有效的点云特征.使用深度霍夫投票算法与均值偏移聚类算法计算实例的三维关键点坐标.最后,利用最小二乘拟合方法计算预测三维关键点的物体位姿参数.在LineMOD数据集和YCB-Video数据集上进行测试,实验结果表明:与六自由度物体位姿估计方法相比,本文模型预测的物体位姿准确率高于其他方法,平均准确率分别达到99.5%和96.9%.网络同时基本满足实时性要求,完成一帧RGB-D图像的多实例物体位姿估计时间需0.06 s.
基金co-supported by the National Natural Science Foundation of China (Grant Nos. 61371134, 61071137)the National Basic Research Program of China (No. 2010CB327900)
文摘The application of high-performance imaging sensors in space-based space surveillance systems makes it possible to recognize space objects and estimate their poses using vision-based methods. In this paper, we proposed a kernel regression-based method for joint multi-view space object recognition and pose estimation. We built a new simulated satellite image dataset named BUAA-SID 1.5 to test our method using different image representations. We evaluated our method for recognition-only tasks, pose estimation-only tasks, and joint recognition and pose estimation tasks. Experimental results show that our method outperforms the state-of-the-arts in space object recognition, and can recognize space objects and estimate their poses effectively and robustly against noise and lighting conditions.
文摘Conventionally, image object recognition and pose estimation are two independent components in machine vision. This paper presented a simple but effective method KNNSNG, which tightly couples these two com ponents within a single algorithm framework. The basic idea of this method came from the bionic pattern recog nition and the manifold ways of perception. Firstly, the shortest neighborhood graphs （SNG） are established for each registered object. SNG can be regarded as a covering and triangulation for a hypersurface on which the training data are distributed. Then for recognition task, the deter mined test image lies on which SNG by employing the parameter ＂k＂, which can be calculated adaptively. Finally, the local linear approximation method is adopted to build a local map between highdimensional image space and lowdimensional manifold for pose estimation. The projective coordinates on manifold can depict the pose of object. Experiment results manifested the effectiveness of the method.
