Improving accuracy of feature matching in visual SLAM using spatial consistency of point features

Many RGB-D SLAM systems employ the Iterative Closest Point (ICP) with RANSAC as standard algorithm to align point features. However, when the noise in data increases or the offsets between frames are large, the results of RANSAC could be unreliable. In order to improve accuracy of trajectory estimation in such scenes, a novel approach is proposed for feature matching using spatial consistency of point features in RGB-D SLAM. By taking advantage of spatial structure information of point features, our approach can extract correspondences between frames more reliably than feature matching approaches with RANSAC. Hence, the accuracy of feature matching increases. Results on open dataset show that this approach can improve accuracy and robustness of visual SLAM.