A Motion-Adaptive Parallax-Based Monocular Visual Odometry

Monocular odometry is the pivotal part of the visual simultaneous localization and mapping (SLAM) system. Since monocular odometry algorithms available now are built on some certain motion assumptions, they behave unstable severe and rapid motion, which easily reduces the number of reliable landmarks and causes more error. For this reason, a monocular visual odometry algorithm based on a parallax-based weight matrix with an adaptive feature matching is proposed in this paper. The adaptive feature matching algorithm can quickly and reliably establish feature matching under various motions, which ensures the real-time performance of the algorithm under fast and violent motion. The parallax-based weight matrix estimation can make high-parallax map points with high precision dominate the pose estimation and other map points more accurate in the optimization, which can enhance the accuracy of the result in real-time map update and optimization. The experiments are done on two public data sets KITTI and EUROC, in which our algorithm shows slightly better accuracy in location under normal conditions and much better robustness under fast motion and severe motion compared with ORB-SLAM.