Point Cloud Plane Segmentation-Based Robust Image Matching for Camera Pose Estimation

The mainstream image matching method for recovering the motion of camera is based on local feature matching, which faces challenges rotation, illumination, and presence dynamic objects. In addition, relies distance between descriptors, easily leads to lots mismatches. this paper, we propose a new robust pose estimation, called IM_CPE. It novel descriptor combined with 3-D point clouds matching. Specifically, extract points pair matched cloud planes, are generated segmented depth images. Then, their corresponding planes descriptors. Moreover, robustness can be guaranteed by centroid planes. We evaluate performance IM_CPE using four well-known key extraction algorithms, namely Scale-Invariant Feature Transform (SIFT), Speed Up Robust (SURF), Features from Accelerated Segment Test (FAST), Oriented FAST Rotated Brief (ORB), sequences TUM RGBD dataset. According experimental results, compared original SIFT, SURF, FAST, ORB NN_mAP algorithms has been improved 11.25%, 13.98%, 16.63%, 10.53% average, respectively, M.Score also 25.15%, 23.05%, 22.28%, 11.05% respectively. results show that existing significantly improve these algorithms.