Research On Visual Navigation Methods Of Mobile Robots In Unknown Indoor Environments

Localization,mapping,and path planning are key issues for mobile robots to achieve autonomous navigation functions.Due to the advantages of low price and rich information of visual sensors,visual simultaneous localization and mapping(visual SLAM)has become one of the research hotspots in recent years.Many classical algorithms,such as Dynamic Window Algorithm(DWA)and Timed Elastic Band(TEB),have appeared in the field of robot path planning research.However,there are still problems such as the inability of mapping quality to meet navigation requirements and the poor path planning effect,when a simple combination of visual SLAM and traditional path planning is applied to navigation tasks in unknown indoor environments.To this end,this thesis proposes a two-dimensional(2D)grid map construction method and a path planning method based on deep reinforcement learning(DRL)under the ORB-SLAM2 framework,which effectively achieves autonomous navigation of mobile robots in complex unknown indoor environments such as low textures and multiple obstacles.The main research work is as follows:Aiming at the problem that the map created by visual SLAM cannot meet the navigation requirements,a 2D grid map for navigation is constructed based on the ORB-SLAM2 algorithm.First,a point cloud map construction thread with key frame filtering strategy and point cloud filtering algorithm is created to realize the real-time construction of dense point cloud maps on the CPU.Then,the ground point cloud is separated,and the point cloud map is converted into an octree map(Octomap).Finally,the Octomap with three-dimensional(3D)space occupancy information is projected to a 2D grid map to realize the construction of the navigation map.To solve the problem that the feature points are easily lost when combining visual SLAM with traditional path planning algorithms,a path planning method based on DRL is proposed.Considering factors such as tracking stability and navigation safety,a multi-sensing network structure and multi-constrained reward function are designed,thereby improving the algorithm’s ability to adapt to the scene and the speed of convergence.The performance of the method is verified in the simulation environment Habitat,and the multi-constraint reward mechanism is compared and analyzed to verify its effectiveness.To verify the navigation effect,a robot experimental platform is built based on the robot operating system(ROS)in real indoor scenes,and point-to-point navigation tasks in three real indoor scenarios are performed.Experimental results show that the algorithm proposed in this thesis performs well in the real scenes.The average navigation success rate in three scenes is about 66.6%,and it performs well in avoiding visual SLAM tracking loss.Compared with traditional path planning methods,the average rate of tracking loss is reduced by 54.8%.