Hierarchical Semantic Mapping And Application For Robot Service Navigation

To provide help for human beings,robots need to complete both navigation and search tasks autonomously in dynamic scenes.Map serves as a carrier for robots to understand the environment and interact with people.Reasonable maps are relied on to carry out service navigation tasks,and the integration of semantic information into maps can effectively improve the ability of robot to finish tasks.Based on the current drawbacks of robot service navigation,such as insufficient understanding of environment,single type of work,inappropriate path generation and so on,the construction of hierarchical semantic map and the corresponding service navigation strategy in the complex indoor dynamic environment is studied systematically,which improves the robot's ability to complete the service navigation task.To obtain semantic information of indoor environment using RGB-D data,a method of overall semantic segmentation is proposed based on environmental feature voxel grid.Firstly,dense point cloud is constructed by orb-slam and is voxelized to establish the overall representation of the environment.The environmental feature voxel grid is generated by mapping the vision and depth detection features.The idea of mask RCNN is extended to 3D.A new 3D region recommendation network is designed using 3D and 2D convolutional neural network and the synchronous object instance segmentation,category and mask prediction are realized.The proposed method is tested on Scan Net dataset and compared with several up-todate 3D semantic segmentation methods to verify the accuracy of the proposed method.To solve the problem that a single map can not meet the demands of service navigation tasks in dynamic indoor environment,a hierarchical semantic map with 3D semantic map as the core layer and global cost map and item distribution map as the application layer is established.The core layer is built by segmenting instances and is stored it in the database.The global cost map is selected for path planning with the added geometric static layer and geometric obstacle layer to Enhance robot semantic perception.In the search part,the instance association model is established using the skip-gram model,and the probability map of instance distribution is established using Gaussian distribution model,which can effectively improve the search efficiency with semantic information.The corresponding maps are established in the simulation and real environment respectively,which verifies the effectiveness of the hierarchical map construction algorithm.Based on the constructed hierarchical semantic map,a complete service navigation strategy is developed and implemented.First,the order is preprocessed to split the mobile tasks of robots.Then the global and local path planning methods are given,and the obstacle detour strategy and path congestion recovery strategy are optimized to improve the safety and intelligence of the robot navigation path.Aiming at the search task,the information gain and path length are used to calculate the item search cost.The new path nodes are generated using the sampling and path reconnection mechanism.By comparing with several common navigation and search strategies,the superiority of service navigation strategy is verified.Rich semantic information is applied in the system to construct hierarchical map,which can efficiently improve the robot's ability to understand the environment and to deal with human-computer interaction tasks flexibly.Each module is closely related to achieve the goal of autonomous service navigation.