基于旋转投影二进制描述符的空间目标位姿估计

发布时间：2018-03-20 23:24

本文选题：空间目标　切入点：位姿估计　出处：《光学精密工程》2017年11期 　论文类型：期刊论文

【摘要】：为了实现基于点云的空间目标相对位姿快速估计,提出一种旋转投影二进制描述符(BRoPH)。该描述符首先建立特征点处的局部参考坐标系,然后通过旋转投影局部点云生成不同视角下的密度图像块和深度图像块,最后根据图像块生成特征点的多尺度二进制字符串。针对位姿估计对实时性的要求,在分析BRoPH Hamming距离分布的基础上,提出了基于Hamming距离阈值的特征匹配策略,用于剔除潜在的错误配对,加快位姿估计收敛速度。最后,在基于局部特征描述符位姿估计框架下分别与SHOT描述符和FPFH描述符进行了比较。结果表明:BRoPH描述符在仅需要SHOT和FPFH平均内存1/80的基础上,得到了远高于SHOT和FPFH的平均位姿估计精度,其平均姿态误差小于0.1°,平均位置误差小于1/180 R。此外,基于Hamming距离阈值的特征匹配策略使得BRoPH的位姿粗估计速度加快了7倍,总体位姿估计频率超过7Hz,比SHOT和FPFH分别快3~6.8倍。该方法具有占用内存小、计算速度快、位姿估计精度高和抗干扰能力强等优点,满足基于点云的空间目标位姿估计实时性要求。
[Abstract]:In order to estimate the relative position and orientation of spatial objects based on point cloud, a rotating projection binary descriptor (BRoPHG) is proposed. The local reference coordinate system at the feature points is established in the descriptor. Then the density image block and depth image block under different angle of view are generated by rotating projection local point cloud. Finally, the multi-scale binary string of feature points is generated according to the image block. Based on the analysis of BRoPH Hamming distance distribution, a feature matching strategy based on Hamming distance threshold is proposed to eliminate potential error pairs and speed up the convergence of pose estimation. Compared with SHOT descriptor and FPFH descriptor under the framework of local feature descriptor position and pose estimation, the result shows that the SHOT and FPFH average memory 1/80 are only needed in the SHOT descriptor. The average attitude error is less than 0. 1 掳and the average position error is less than 1/180 R. in addition, the feature matching strategy based on Hamming distance threshold accelerates the coarse estimation speed of BRoPH position and pose by 7 times. The total pose estimation frequency is more than 7 Hz, which is 36.8 times faster than SHOT and FPFH, respectively. This method has the advantages of small memory, fast computing speed, high precision of pose estimation and strong anti-interference ability. It can meet the real-time requirements of spatial target pose estimation based on point cloud.
【作者单位】：清华大学自动化系;清华大学深圳研究生院;
【基金】：广东省自然科学基金资助项目(No.2014A030310318,No.2015A030313881) 深圳基础研究资金资助项目(No.JCYJ20140509172959962,No.JCYJ20160301153317415,No.JCYJ20160301100921349)
【分类号】：TP391.41

【相似文献】