Research On Robot Navigation And Positioning Technology Based On Multi-sensor Fusion

In order to realize the autonomous navigation of mobile robots,considering the complexity of the indoor environment,this dissertation conducts monocular depth information research,multi-sensor fusion SLAM(Simultaneous Localization and Mapping)algorithm and path planning for depth information acquisition in the indoor environment.research.In this dissertation,a method of depth information estimation based on monocular depth estimation is proposed to solve the problem that map information is lost during the process of using visual SLAM to build a map,which leads to insufficient mapping accuracy and cannot be used for navigation.Using ORB SLAM3(Orinted FAST and Brief The RGBD mode in Simultaneous Localization and Mapping3)is used to construct dense maps to effectively improve the accuracy of maps and complete complex indoor scene navigation and positioning.The main research contents of this dissertation are as follows:First of all,in view of the low accuracy of single-sensor SLAM mapping,the high cost of different multi-sensor SLAM mapping and the susceptibility to environmental interference,this dissertation compares the advantages and disadvantages of different sensors and different multisensor fusion methods,and chooses to use Deep learning estimates depth information,inertial sensor(IMU,Inertial Measurement Unit)and multisensor fusion method of monocular camera to perform multi-sensor fusion SLAM based on ORB-SLAM3 to construct dense point cloud map to improve mapping accuracy.Then,aiming at the problem of blurred edges in the supervised monocular depth estimation algorithm,a multi-scale fusion monocular depth estimation algorithm based on the pyramid vision transformer2 encoder and lightweight decoder is proposed.By adaptively selecting multi-scale key features for fusion,the algorithm can improve the problem of object edge blur,and by comparing with CNN-based depth estimation algorithm experiments,the algorithm can shorten the training time by reducing the number of training layers of the neural network.Secondly,in view of the problems of expensive production cost and cumbersome production steps of labeled datasets in supervised monocular depth estimation,at the same time,the unsupervised monocular depth estimation algorithm has the problem of insufficient accuracy and large errors.In this dissertation,we propose a U-network with multi-scale learnable CBAM attention mechanism.The network can effectively extract the features extracted by the pose and depth estimation network and fully integrate different scales,which can effectively reduce the error,and in order to reduce the influence of the depth information obtained by the dynamic object,use the mask to remove the depth information of the dynamic object.Finally,build a multi-sensor fusion mobile robot hardware platform,make a color picture dataset for indoor complex environments,and perform depth estimation on the constructed dataset to verify the depth estimation algorithm proposed in this dissertation,and then input the obtained depth information and color pictures into SLAM mapping is performed in the system.Finally,the constructed dense point cloud map is converted into a two-dimensional grid map,and on this basis,the autonomous navigation of the mobile robot is realized to complete the navigation and positioning experiment of the mobile robot.