Research And Implementation Of Multi-robot SLAM Algorithm Based On Vision

In the era of rapid development of robotics technology,the application of multi-robot systems in practical life has received increasing attention.Their superior flexibility,robustness,and efficiency enable robots to better accomplish tasks.Among them,Multi-Robot Simultaneous Localization and Mapping(Multi-Robot SLAM)has become a hot research project for researchers,which refers to multiple robots running simultaneously in unknown environments to achieve real-time localization and map construction.In order to improve the localization and mapping accuracy of robots in complex environments,this thesis uses a depth camera and an Inertial Measurement Unit(IMU)as sensors to build a multi-robot collaborative localization and mapping system,with the main work as follows:In response to the problem that the front-end of pure visual multi-robot collaborative localization and mapping cannot effectively complete navigation positioning and map construction in complex environments or poor image quality,this thesis improves the visual odometry module by fusing visual data and inertial data to improve the accuracy and robustness of the front-end.First,a loosely coupled method is used to estimate the initial state of the visual-inertial odometry,and then the pose information between adjacent keyframes and pre-integrated measurement data are aligned within a discrete time interval using the local bundle adjustment method.Next,the Allan variance is used to eliminate the random errors of the IMU,improving the response speed and accuracy of the measurement data.Finally,the visual-inertial data fusion is achieved based on graph optimization.The root mean square error of the localization test in this thesis ’s algorithm is 0.1228 m,and the overall performance is better than the commonly used multi-robot collaborative localization algorithm CCM-SLAM.To address the problem of data redundancy in multi-robot collaborative mapping,this thesis proposes a keyframe selection strategy based on disparity and tracking quality.The current environmental information is used for feature extraction,and rich and motion representative scenes are selected as keyframes in the continuous image sequence,and images that do not meet the keyframe conditions are all deleted to relieve the communication pressure caused by limited bandwidth resources,improving the system’s computational efficiency and real-time performance.Compared with CCM-SLAM,the running time of this thesis’ s algorithm is reduced by about 17%.In response to the fact that most current multi-robot collaborative mapping builds sparse point cloud maps with low accuracy and difficulty expressing environmental details,this thesis designs and implements a multi-robot dense map that can be used in actual scenes.The sub-end collects point cloud data and transmits it to the central end through a communication module to construct a point cloud map,and real-time 3D octree maps and2 D occupancy grid maps are constructed,making the system applicable to navigation work.