Robust dense optical flow with uncertainty for monocular pose-graph SLAM

In this paper, we propose how to use dense optical ow field as opposed to sparse feature matches to improve large-displacement monocular visual odometry. The principled framework we developed incorporates uncertainties in the construction of a four-dimensional cost volume for dense ow computation. A novel weighted eight-point algorithm is derived which robustly estimates inter-frame camera motions by using the obtained dense correspondences with uncertainties. This initial motion estimation is subsequently employed to achieve potential loop closing operation, optimised jointly in a robust pose-graph SLAM framework. Performance of the proposed new method has been validated on standard benchmark dataset - KITTI dataset. Experimental results demonstrate that the accuracy of our method is on par with other state-of-the-art methods without relying on commonly used priors such as motion constraint or ground plane segmentation.