Multi-robot Lifelong Learning Place Recognition Method Based On Semantic Constellation

Place recognition is the key technology to solve the problem of mobile robot confirming its position,and has always been a research hotspot in the field of robot.With the wide application of various sensors in mobile robots,place recognition in complex scenes is a challenging problem that has not been solved for a long time.The challenges mainly come from three aspects: first,the change of image angle caused by different camera shooting angles;second,the change of image appearance caused by drastic changes in light and weather;third,the fault tolerance rate is low in a wide range of operating environment.Aiming at the above problems,this paper proposes a multi-robot place recognition method based on semantic constellation.Firstly,aiming at the problem that the observation from different angles affects the recognition accuracy,a multi-robot place recognition method based on semantic constellation is proposed.When the robot has explored a scene,it identifies the objects in the scene,constructs the semantic constellation according to its spatial relative position and semantic information,and extracts the semantic descriptor of the observation object in the space.When viewed from different angles,the same semantic constellation can be recognized as the same place,and transmits the compact semantic descriptor to each robot to determine the number of other robots who have observed the scene,Then the complete semantic constellation is sent to the matching robot to verify the matching using geometric information.Experiments show that the proposed algorithm improves the recognition accuracy of the system from different angles and reduces the memory required for data packets between machines.Secondly,aiming at the influence of the semantic change of cloud and illumination on the accuracy of darg image,a new method is proposed.In this paper,segnet network is used to generate and train the semantic segmentation map in RGB image,quantify some feature points in the point cloud,represent them with three-dimensional voxels,represent the feature points in three-dimensional space,and output their coordinates.Through geometric consistency matching,all relevant points in the point cloud are mapped to RGB image,so as to combine visual feature information with Li DAR point cloud information.Through CARLA simulator,the urban experimental environment under various climatic conditions can be generated by ourselves.Experiments show that the combination of RGB image information and Li DAR point cloud descriptor improves the recognition accuracy of feature points in 3D space,and improves the robustness to changes in lighting,weather and other conditions.Finally,aiming at the low fault tolerance of the system in a wide range of operating environment,a multi task learning SLAM algorithm based on lifelong learning is proposed.For the problem that the system occupies too much memory in the long-period environment,this paper adopts a SLAM algorithm based on multi-part map framework to update the sub map in time,and adopts the methods of position and attitude map thinning and sub map pruning based on Chow Liu tree.For the problem of catastrophic forgetting of feature objects in the process of place recognition,this paper adopts the multi task allocation method based on lifelong learning to maintain a sparse and shared basis for all task models,transfer the previously observed place as the basic knowledge to each new task model,and redefine the basis with the passage of time,so as to maximize the performance of all tasks.Experiments show that the proposed algorithm can greatly reduce the utilization of memory and CPU resources,avoid the problem of catastrophic forgetting,and improve the accuracy of the system.