Tracking by Detection: Robust Indoor RGB-D Odometry Leveraging Key Local Manhattan World

In indoor scenes, pose estimation drift in SLAM systems is challenging to mitigate due to factors like low texture and texture repetition. Several studies utilize Manhattan structural information to achieve low-drift rotation estimation consequently reducing the cumulative error. However, the application scenarios of these methods are limited by the Manhattan assumption. To address this problem, we propose a robust RGBD odometry for tracking in more general structured scenes, which can represent a wider range of scenes by leveraging Atlanta World assumption, consisting of a vertical direction and multiple horizontal directions. To this end, we design tracking by detection algorithms for extracting Key Local Manhattan World in each frame, which is defined as the most stable structural information in current frame. Specifically, we detect local Manhattan Worlds in current frame, if a more stable structural feature emerges, it becomes the new tracking target. The angle between these Manhattan Worlds are then computed, constructing the Atlanta World. For accurate pose estimation, we use Key Local Manhattan World to first estimate low-drift rotation followed by estimating translation using point-line structural features. Extensive experiments on public benchmark and self-recorded datasets show that our method outperforms existing state-of-the-art methods with a significant margin of 22% while extending their applicability to the Atlanta World.