RGBDTAM: A cost-effective and accurate RGB-D tracking and mapping system

Abstract

Abstract-Simultaneous Localization and Mapping using RGB-D cameras has been a fertile research topic in the latest decade, due to the suitability of such sensors for indoor robotics. In this paper we propose a direct RGB-D SLAM algorithm with state-of-the-art accuracy and robustness at a los cost. Our experiments in the RGB-D TUM dataset [34] effectively show a better accuracy and robustness in CPU real time than direct RGB-D SLAM systems that make use of the GPU.The key ingredients of our approach are mainly two. Firstly, the combination of a semi-dense photometric and dense geometric error for the pose tracking (see Figure 1), which we demonstrate to be the most accurate alternative. And secondly, a model of the multi-view constraints and their errors in the mapping and tracking threads, which adds extra information over other approaches. We release the open-source implementation of our approach 1 . The reader is referred to a video with our results 2 for a more illustrative visualization of its performance.

Dense tracking, mapping and scene labeling using a depth camera

“…For the experiment with the TUM RGB-D dataset, we got 5.2cm, 8.3cm and 3.6cm for translation RMSE. The accuracy of our system is on par with [5] but lower than systems such as [12][13][14][15] that integrate techniques for achieving global consistency such as appearance-based place recognition for loop closure detection, non-rigid space deformation, pose-graph optimization and bundle adjustment. Second, the system updates the color of the model every frame, getting a 3D model with color coming from the scene or with color coming from the object detection module.…”

“…Place recognition, based on bag of words, is used for relocalization in case of tracking failure. RGBDTAM [15] combines a semi-dense photometric error (only for pixels belonging to edges) and a dense geometric error (point cloud alignment in a coarse-to-fine scheme with a pyramid of four levels) for camera tracking, achieving CPU real time performance. The tracking thread minimizes the geometric and photometric reprojection error with respect to a previous keyframe.…”

“…Other difference with [5] is the color management; we integrate color information sequentially during the reconstruction process similar to Kintinuous [6], but unlike it, our system updates color information of the observed voxels by averaging the accumulated color data with the current one, in a similar way to the update of the TSDF value. Moreover, our system does not use RGB data for tracking (only for coloring the model) and does not perform loop closures with techniques such as space deformation, pose-graph optimization or bundle adjustment, like in [12][13][14][15], that makes them more robust to long and complex trajectories. We do not track and reconstruct moving objects like [19], [20], but we detect and reconstruct a static desktop-like scene.…”

“…It is due to similarities in the design: dense ICP for tracking and volumetric structure with TSDF values for the model. More recent systems [12][13][14][15], use both photometric and geometric error for estimating the camera pose. Moreover, they include modules for achieving loop closure detection and global consistency with techniques such as non-rigid deformation, pose-graph optimization and bundle adjustment.…”

Globally Consistent Indoor Mapping via a Decoupling Rotation and Translation Algorithm Applied to RGB-D Camera Output

“…However, it also uses measurements from depth images directly, and the uncertainty model of the depth measurements was not well studied. The work in [31] combined measurements from the depth and RGB images, and is suitable for large environments. The Kinect fusion reduced the noise in the depth image by ray casting a synthetic depth image from a previous model [11].…”