Robust direct vision-based pose tracking using normalized mutual information

This paper presents a novel visual tracking approach that combines the NMI metric and the traditional SSD metric within a gradient-based optimization frame, which can be used for direct visual odometry and SLAM. We firstly derivate the closed form expression for first- and second-order analytical NMI derivatives under the assumption of rigid-body transformations, which then can be used by subsequent Newton-like optimization methods. Then we develop a robust tracking scheme that utilizes the robustness of NMI metric while keeping the optimization characteristics of SSD-based Lucas-Kanade (LK) tracking methods. To validate the robustness and accuracy of the proposed approach, several experiments are performed on synthetic datasets as well as real image datasets. The experimental results demonstrate that our approach can provide fast, accurate pose estimation and obtain better tracking performance over standard SSD-based methods in most cases. © 2018 SPIE.