Abstract: A robust visual odometry in the framework of deep self-supervision, which can overcome the interference of dynamic regions in camera pose estimation and scene depth computation, was proposed. Two consecutive frames of images were selected. Depth estimation network was first used to calculate the depth maps of the two images, and pose estimation network was used to calculate the relative poses. Then, dynamic regions were obtained by comparing a depth map with the one synthesized by warping the other depth map into the current view. Based on the detected dynamic regions, the dynamic and static parts in input images were seperated with this algorithm. Then, the features in the static regions of the two images were matched to obtain the final camera poses. The loss function was composed of photometric error, smoothness error, and geometric consistency error, and dynamic region information was integrated into the loss function. Finally, this loss function was used to train the proposed network in a self-supervised end-to-end manner. Experiments were carried out on the widely used KITTI dataset. Specifically, the proposed model was compared with the state-of-the-art ones proposed in recent two years. Experimental results show that this algorithm can deal with the dynamic scenes more robustly and achieve higher accuracy in camera pose estimation and scene depth computation.