基于视觉与惯性组合信息的图像特征提取与匹配

Image feature extraction and matching based on information from visual and inertial sensors

在采用视觉与惯性组合方式进行导航定位时,针对序列图像帧特征提取与匹配过程中存在误匹配及剔除算法耗时长的问题,提出基于两点RANSAC算法的误匹配剔除改进算法。首先采用交叉验证法进行粗剔除以保证特征点匹配的唯一性,并降低外点比例获得粗匹配结果;然后结合惯性传感器中的陀螺仪信息,使用两点RANSAC算法进行精剔除;最后运用精剔除获得的匹配结果重新估计本质矩阵,获得满足该本质矩阵的内点集。实验结果表明:与传统RANSAC算法相比,该算法能获得更好的匹配结果,耗时只有传统RANSAC算法的1/10。

Using the combination of visual and inertial methods for navigation and positioning,the sequence image frames from the vision sensor are obtained and need to be matched.However,it takes a long time to eliminate the mismatch feature points using the traditional image feature extraction and matching algorithm.An improved mismatch points rejection method based on two-points RANdom Sample Consensus(RANSAC)algorithm is proposed.The Cross-matching filter algorithm is adopt to ensure the matching uniqueness of feature points and produce a coarse matching result,a refinement matching process is carried out by using a two-points RANSAC algorithm with the combination of inertia sensor gyroscope information.Finally,the essential matrix of the matching result is re-estimated according to the inner points obtained by the fine elimination,and on the contrary,new collection of inner points is further optimized by the re-estimated essential matrix.The experimental results show that,compared to the traditional algorithm of RANSAC,the proposed method can produce better matching result and be 10 times faster.