Multi-robot Visual Laser-SLAM In Large-scale Environment

Simultaneous localization and mapping is the key technology for mobile robots,especially for the robots working in the unknown environments.As the robot technology grows rapidly,many algorithms of SLAM are proposed;however,most approaches were designed for single robot.But for applications in large-scale or complicated environments,it is difficult to utilize a single robot to complete satisfied performance.Therefore,this dissertation proposes a multi-robot SLAM based on laser and visual information,solving the disadvantages of single robot in aspect of the efficiency,accuracy and robustness.There are several challenges for multi-robot SLAM.Firstly,the robots have no prior knowledge about the relative pose between different robots,which make it difficult to establish links between different robots.Second,the map fusion between robots will seriously be affected by the cumulative error in each robot.Third,wrong constraint detections between robots can ruin the whole map.To address the challenges,this dissertation develops a new SLAM approach with a distributed robot system based on laser and visual information.Laser SLAM has a good performance in small environments.However,when the robot works in a complicated or large-scale environment,it is easy to detect wrong loops,leading to the SLAM failure.In contrast,the camera can provide rich information to recognize the place,and therefore,the visual information is employed to aid the laser SLAM,with the purpose to detect the loop closure and eliminate the cumulative error of single-robot.Furthermore,multiple robots exchange their visual information through TCP/IP Socket,and the visual information can also be used to establish the links between different robots.When a robot visits the place that another one has visited,the constraints between different robots can be established through the proposed place recognition algorithm.The constraints establish a bridge to fuse the sub-map of each robot.With these constraints,the graph nodes of robots can be transformed to a master robot for generating a fused map.The place recognition algorithm is not only adopted in the loop closure detection of the single-robot,but also in the process of building the constraints between different robots.The ORB features and bags of words are adapted to provide real-time performance.In this dissertation,several experiments were performed to prove the effectiveness of the proposed approach.The generated maps by the proposed approach is consistent with the real environments,which verifies the effectiveness of the algorithm.The proposed approach can be widely used in intelligent service robots,autonomous guided vehicles,unmanned aerial vehicles and other fields,and thus has a great significance for the development of the robotics in China.