Semantic Map Construction Based On SLAM Algorithm And Convolutional Neural Network

The map is a model of the task environment in which the robot is located,including the features in the environment.With the development of mobile robots in a more automated and intelligent direction,the shortcomings of traditional maps,such as less environmental information and less distinguishing of environmental features,have become increasingly apparent.The semantic map is a multi-level map that contains the semantic information of the environment or the semantic information of the objects in the environment.It contains rich environment information and can help the robot better understand the environment.On the basis of the traditional 2D grid map,semantic labels of target objects are added to design and implement the semantic map based on SLAM algorithm and convolutional neural network.In this thesis,Turtlebot2 is used as the hardware platform of mobile robot,ROS is used as the robot software operating system to complete the following parts of work.(1)Multi-sensor data fusion.In view of the fact that there is a certain system error in the odometer and the slipping and idling may occur in the actual movement,the data of the IMU and the odometer are integrated to obtain more accurate odometer data.Also,in view of the situation that the lack of environmental information obtained by2 D lidar,the data of the lidar and the depth camera are merged to obtain richer spatial environmental information.(2)Autonomous exploration of mobile robots.Aiming at the situation that mobile robots need to perform tasks individually,the boundary-based exploration algorithm is used as the autonomous exploration algorithm,and the dynamic window algorithm is used as the local path planning algorithm to realize the autonomous exploration of the mobile robot.With the Cartographer algorithm,2D grid map of the environment can be built independently.(3)Target object recognition and semantic information addition based on lightweight convolutional neural network.In view of the lack of computing power of mobile robot,a lightweight convolutional neural network is designed based on the network pruning method to achieve target recognition.Based on the recognition results of the lightweight convolutional neural network and the 2D grid map,Marker plugin is adopted to add the semantic labels of the target object and complete the semantic map construction.The experimental results show that the positioning accuracy of the mobile robot is improved after the integration of IMU and odometer data.By fusing the data of lidar and depth camera,the mobile robot can detect the obstacles at different heights and obtain more abundant space environment information.Based on autonomous exploration algorithm and SLAM algorithm,mobile robot can construct 2D raster map of indoor environment independently.By using lightweight convolutional neural network,there still have high object recognition accuracy even when the number of parameters and model size are greatly reduced.Combined with the mobile robot autonomous exploration algorithm and SLAM algorithm,the independent construction of indoor environment two-dimensional semantic map is realized.The system can extract the semantic information in the environment,and include the semantic features of the environment on the basis of the grid map,deepen the robot's understanding of the environment,and have a better promotion of human-computer interaction.