Sparse edge visual odometry using an RGB-D camera

We propose a sparse visual odometry model for RGB-D cameras, which minimizes the photometric errors over dispersed edge points. In contrast to the feature-based methods, we use cells to extract the features on the edge images. This allows us to maintain the robustness of the information and make the computation more efficient. Furthermore, the different degree of exposure is represented as a posterior probability in each feature points. We can adjust the weights to improve the pose according to the probability. Since the estimate might not be accurate due to the feature points affected by sensor noise, we use the geometry and mixture distribution to update the depth values. The PnP algorithm is then used to adjust the pose again and reduce the camera drift in the frontend process. Experiments are carried out using public datasets to demonstrate the effectiveness of the proposed method.