Efficient RGB–D data processing for feature–based self–localization of mobile robots

Abstract

The problem of position and orientation estimation for an active vision sensor that moves with respect to the full six degrees of freedom is considered. The proposed approach is based on point features extracted from RGB-D data. This work focuses on efficient point feature extraction algorithms and on methods for the management of a set of features in a single RGB-D data frame. While the fast, RGB-D-based visual odometry system described in this paper builds upon our previous results as to the general architecture, the important novel elements introduced here are aimed at improving the precision and robustness of the motion estimate computed from the matching point features of two RGB-D frames. Moreover, we demonstrate that the visual odometry system can serve as the front-end for a pose-based simultaneous localization and mapping solution. The proposed solutions are tested on publicly available data sets to ensure that the results are scientifically verifiable. The experimental results demonstrate gains due to the improved feature extraction and management mechanisms, whereas the performance of the whole navigation system compares favorably to results known from the literature.

Recent Advances in 3D Data Acquisition and Processing by Time-of-Flight Camera

“…2. The depth calculation (4) assumes that the IR signal comes from a specific 3D point in the scene during the IT. If there is any motion in the period, the resulted depth will be corrupted.…”

“…This technology has been applied in many fields for research and engineering solutions. Some practical applications of this sensing modality include robot navigation [3], [4], collision and obstacle detection for robot-assisted surgery [5], 3D reconstruction [6], measurement of structural deformation [7], [8], simultaneous localization and mapping (SLAM) [9], [10], human-computer interaction [11], 3D television (3DTV) [12], plant phenotype [13], [14], debris monitoring [15], etc.…”

On the Application of RGB-D SLAM Systems for Practical Localization of Mobile Robots

“…From the RGB images, it only extracts the keypoints with the use of the ORB detector [26], which was chosen due to the good trade-off between the performance in visual navigation and the computational ef iciency [28]. The keypoints are extracted with respect to the scene depth data availability at the given point, avoiding artifacts in depth images [18]. In comparison with our earlier localization systems [3,18], an entirely new SLAM architecture has been introduced in [4].…”

“…The keypoints are extracted with respect to the scene depth data availability at the given point, avoiding artifacts in depth images [18]. In comparison with our earlier localization systems [3,18], an entirely new SLAM architecture has been introduced in [4]. The typical approach to graph-based SLAM, which dates back to 2D SLAM systems employing laser scanners [13], is based on the optimization of a graph of sensor poses and explicit detection of loop closures.…”

“…The most important of them is the use of a computationally ef icient VO pipeline for the frame-to-frame tracking of the sensor pose. This subsystem is based on the LucasKanade algorithm [1] and uses only the RGB images [18]. This solution allows to better estimate the sensor pose with respect to the map than a camera motion model applied in most of the visual SLAM systems [22].…”