Visual Odometry integrated with Self-Supervised Monocular Depth Estimation

Abstract

To solve the problem of poor robustness of multiview geometric visual odometry in a dynamic environment, we propose a visual odometry method base on deep learning that combines depth estimation and geometric pose determination. First, we remove the time-consuming dense optical flow prediction network in GeoNet. We propose a method for image reconstruction at one frame interval to solve large displacement between moving objects. Second, we propose a way to find the best pixel to solve the problem of pixel occlusion in the image. Third, to further solve the uneven illumination of the image and the degradation of image quality, we propose an adaptive histogram equalization based on limited contrast to enhance the image. A large number of experimental demonstrations have been carried out on the KITTI public dataset. The experimental results show that our network reduces the time cost and complexity on the general evaluation index. It has a significant improvement compared with GeoNet and has achieved more accurate in-depth and position prediction results.