复杂场景下基于C-SHOT特征的3D物体识别与位姿估计

发布时间：2018-04-30 15:43

本文选题：三维物体识别 + 位姿估计　；参考：《计算机辅助设计与图形学学报》2017年05期

【摘要】：为了准确地同时识别复杂点云中的多个目标,提出一种基于法矢改进点云特征C-SHOT的3D物体识别方法.首先,在估计RGB-D数据的点云法矢时将邻域点距离信息考虑在内,计算带距离权重的协方差矩阵得到更精确的点云法矢;其次根据特征点处法矢与邻域法矢的夹角余弦构造点云形状直方图,同时统计点云纹理直方图并与形状直方图融合成C-SHOT描述符;最后对场景与模板分别提取C-SHOT特征,利用Kd树快速求得对应对,引入3D霍夫投票机制,并结合点云局部坐标系克服噪声遮挡问题完成多目标初识别.基于LM-ICP实现精确定位及位姿估计,画出目标包围盒,采用基准数据库CVLab以及采集实验室真实数据进行实验,结果验证了该方法的有效性与精确性.
[Abstract]:In order to accurately identify multiple targets in complex point clouds simultaneously, a 3D object recognition method based on normal vector improved point cloud feature C-SHOT is proposed. Firstly, when estimating the normal vector of the point cloud of RGB-D data, the neighborhood distance information is taken into account, and a more accurate point cloud normal vector is obtained by calculating the covariance matrix with distance weight. Secondly, the point cloud shape histogram is constructed according to the angle cosine between the normal vector and the neighborhood normal vector at the feature point, and the point cloud texture histogram is counted and fused with the shape histogram to form the C-SHOT descriptor. Finally, the C-SHOT feature is extracted from the scene and template, respectively. By using KD tree, 3D Hough voting mechanism is introduced, and the local coordinate system of point cloud is used to overcome the problem of noise occlusion to achieve multi-target initial recognition. The accurate location and pose estimation are realized based on LM-ICP, the target bounding box is drawn, the benchmark database CVLab is used and the real data of the laboratory are collected. The results show that the method is effective and accurate.
【作者单位】：中国民航大学智能信号与图像处理天津市重点实验室;
【基金】：国家自然科学基金(61179045)
【分类号】：TP391.41

【相似文献】