基于立体视觉与特征匹配的点云目标识别算法

Point cloud target recognition algorithm based on stereo vision and feature matching

准确提取三维点云数据中待测目标的点云集合是三维点云目标识别技术的一个关键问题,也是近年来目标识别领域从二维向三维拓展的一个重要挑战,其主要难点在于快速寻找离散点云之间的相关函数关系。结合立体视觉与特征匹配构建了可以表征不同视场条件下的目标点云约束的机制,通过采用立体视觉作为约束条件完成了对原有特征匹配算法的优化。设计了基于立体视觉的估计算法,通过训练学习获得了不同选取比例条件下的识别规则。实验采用ARIES激光雷达采集点云,并通过MATLAB选取三种典型目标状态。当目标区分度高时,优化前后的目标识别率都在98%以上;当目标区分度低时,优化后对目标边界的限定条件可以很好地提高识别概率。采用优化的点云数据位置偏差量可达到0.55 mm,相比未优化的0.74 mm提高了0.19 mm。同时,优化后算法的收敛时间曲线要优于未优化的,3 000点以上的收敛时间均值约为8.33 s,优于未优化的12.76 s。综上所述,优化后的算法具有更好的识别效率。

Accurately extracting the point cloud collection of the target to be measured in the three-dimensional point cloud data is a key issue of the three-dimensional point cloud target recognition technology, and it is also an important challenge in the field of target recognition from 2D to 3D in recent years. The main difficulty is to quickly find the correlation function relationship between discrete point clouds. Combining stereo vision and feature matching, a constraint mechanism of target point cloud that can characterize different field of view conditions is constructed, and the original feature matching algorithm is optimized by using stereo vision as the constraint condition. An estimation algorithm based on stereo vision is designed, and recognition rules under different selection ratios are obtained through training and learning. The experiment uses ARIES Lidar to collect point clouds, and selects three typical target states through MATLAB. When the target discrimination is high, the target recognition rate before and after optimization is above 98%. When the target discrimination is low, the restriction conditions of the target boundary after optimization can improve the recognition probability. The position deviation of the optimized point cloud data can reach 0.55 mm, which is 0.19 mm higher than 0.74 mm before optimization. At the same time, the convergence time curve of the optimized algorithm is better than before optimization. The average convergence time above 3 000 points is about 8.33 s, which is better than 12.76 s before optimization. In summary, the optimized algorithm has better recognition efficiency.