Scale Estimation for Monocular Visual Odometry Using Reliable Camera Height

The development of monocular visual simultaneous localization and mapping (VSLAM) has slowly begun in recent years. At present, the sensors used for VSLAM include monocular, binocular, or depth. For visual mapping, two problems will be encountered and the mapping cannot be performed. One is when there are not enough feature points, the camera's pose at the next moment cannot be estimated, such as a wall. The other is the VSLAM of dynamic environment changes may not be recognized as the same object because the feature matching needs to be coded through its surrounding environment, so it is easy to lose track when encountering changes in light. Compared with binocular and depth, monocular vision lacks depth information, but because it is cheap and easy to install, it needs to be used by multiple people. The current research proposes adding other information, including the camera height, the scene of the reference object and depth estimation by learning methods. The study uses the OpenVSLAM architecture to estimate the camera height scale, and proposes the mechanism be based on the change in the average scale of the first five key frames, with the average scale being updated at the same time to correct the current scale. Through this method, the drastic changes in scale are corrected, and the accuracy of trajectory positioning is improved. We also evaluate our proposed method on a real KITTI Dataset and demonstrate the proposed algorithm is effective and feasible for monocular visual SLAM.