Scale-aware Semi-Direct Monocular Visual Odometry with Points and Lines

Most monocular visual odometry algorithms based on point features have achieved remarkable performance in recent years. However, the existing monocular visual odometry systems have an inherent limitation that the absolute scale cannot be restored. And compared with point features, lines provide far richer information about the structure of the environment. In city or in the so-called “Manhattan worlds” we can obtain abundant line features as a complement to point features. This paper extends PL-SVO [1], a system that combines both points and line segments, by introducing CNN-based monocular depth prediction module. We add this module to the mapping thread in PL-SVO, overcoming the scale ambiguity limitation of the monocular visual odometry. CNN-predicted depth map providing strong prior is employed to initialize the mean and the variance of depth filters corresponding to point and line features. The proposed algorithm called CNN-PL-SVO has been thoroughly evaluated through both popular public datasets (such as KITTI and Karlsruhe) and real-world experiments on our mobile robot platform. The comparison results with state-of the-art monocular approaches (ORB-SLAM and DSO) reveal that our method achieves better performance, in terms of scale consistency and Sim (3) estimation accuracy.