RGB-D Visual Odometry via Low Order Gaussian Gradient Metrics

This paper gives a direct Visual Odometry (VO) of RGB-D cameras by using features conducted with low-order  Gaussian derivative functions such as Gaussian gradient operator. By using the feature metrics locally, it improves further possibilities for sampling more reliable points from scenarios that are lack of structure or texture and is beneficial to continuous tracking. The proposed approach reaches relatively acceptable performance with a globally heuristic framework built on the general coarse-to-fine and inverse compositional estimation. Experimental results are conducted on a group of TUM datasets for validating the proposed approach.