Indoor Slam For Robots Based On Laser And Mono-vision Fusion

SLAM(Simultaneous Localization and Mapping) is one of the most important topics in the robotics field. The existing SLAM frameworks can mainly be divided into laser based SLAM and visual based SLAM(vSLAM).In this dissertation, the laser SLAM and monocular v SLAM methods were firstly carefully studied and analyzed on the advantages and disadvantages respectively. Our experiments also demonstrate that the laser based SLAM is accurate and fast with good performance in small environments. However, due to the accumulative motion error and low density of sensor information, the laser SLAM cannot perform well in complex or large-scale environments. In contrast, the camera can provide very rich measurement information, and thus the vSLAM performs well in environmental identification; however, this method exhibits less accurate mapping performance. And further the vSLAM needs to deal with much more information, which lead to poor real-time performance. In this dissertation, the two frameworks were widely investigated, and then a new approach was designed by employing the monocular visual information to aid the laser SLAM. The advantages of both methods can be blended into one procedure. In this work, the popular ORB features were adopted to build bags of words in order t o improve the operating efficiency. Firstly, we employed visual ORB features to construct environmental bags to achieve the keyframe detection, initial closed-loop detection and global relocalization. These features can help to provide a visual closed-loop signal. And then, the signal triggered the laser SLAM to perform the loop-closure detection, error matching and the pose graph optimization of the robot.Several experiments were performed in the real environments, i.e., a rectangular corridor(26 m×47 m), and a large library(30 m×100 m). From the experimental comparisons, the performance under the proposed approach was improved greatly comparing to the original SLAM methods. Moreover, the approach still exhibited good performance on the computation cost even adding visual processing. The proposed approach in this dissertation can be widely used in service robots, autonomous guided vehicles, unmanned aerial vehicles and so on. It has great significance for the popularization of the robotic applications in China.