A Loop Closure Detection SLAM System by Encoding Visual and Point Cloud Features

摘要

Loop closure detection is a vital aspect of simultaneous localization and mapping (SLAM) and plays a very crucial role in minimizing cumulative mistakes. The hostile and unpredictable environment in which the robot operates makes it unfortunate that the original LIO-SAM technique for loop closure identification utilizing a single sensor is not always accurate. In this work, we describe how vision cameras and LiDAR can be merged. A novel global descriptor, visual scanning context (VSC), is proposed based on the fused data, which retains the geometric properties of the 3D spatial structure described by the visual features and the LiDAR information, and does not rely on training and odometry results, and can be used efficiently for location recognition. To improve the efficiency of retrieving candidate locations, an efficient two-layer search technique is suggested that is viewpoint-independent and rotationally invariant. The performance of our technology is tested and evaluated on publicly available datasets, and trials reveal that our system surpasses previous methods in terms of efficiency, accuracy, and robustness.