Novel Approaches for Periodic Depth Enhancement in Visual SLAM

Most visual simultaneous localization and mapping (SLAM) systems use either monocular or stereo sensor information. However, in some situations, depth information may not be available for each camera frame, for example, when two cameras with different frame rates are used or the computer used does not have the computational resources to calculate a depth map for each camera frame. This work presents two novel approaches by enhancing either each n(k)-th keyframe or each n(f)-th camera frame periodically with depth information. The experimental results on the KITTI visual odometry benchmark show that both approaches improve scale drift compared to monocular SLAM. In terms of trajectory accuracy, the periodic camera frame enhancement outperforms the other novel approach. Even if only a small number of camera frames is enhanced with depth information (10% to 20%), this approach achieves comparable or more accurate results than stereo trajectories. Therefore, this work introduces novel approaches for depth enhancement in visual SLAM systems that do not require depth information at every camera frame. This allows systems with limited access to measured depth or limited computational resources to improve on scale drift.