Keypoint Heatmap Guided Self-Supervised Monocular Visual Odometry

Visual odometry is an important part of visual simultaneous localization and mapping (SLAM) system. In recent years, with the development of deep learning technique, the combination of visual odometry with deep learning has attracted more and more researchers’ attentions. Existing deep learning-based monocular visual odometry methods include a large number of calculations of redundant pixels, and they only consider the pose transformation between two adjacent frames, resulting in error accumulations. To solve the above problems, an end-to-end self-supervised monocular visual odometry method based on keypoint heatmap guidance is proposed in this paper. In the process of network training, the keypoint heatmap is used to guide network learning to reduce the influence of redundant pixels. The photometric error based on the pose consistency constraint of image sequence is calculated to reduce the accumulated error in the pose estimation of video sequence. Extensive experimental results on the KITTI visual odometry dataset have fully validated the effectiveness of the proposed method.