Design And Implementation Of Uav Semantic SLAM Prototypical System Based On RGBD

With the rapid development of computer vision and drone-related technologies,the construction of intelligent drone systems has become a hot topic in the field of robot research,and efficient semantic SLAM(Simultaneously Localization And Mapping)is the basis of advanced tasks such as autonomous positioning,environmental navigation and intelligent search in an unknown environment for the intelligent UAV(unmanned Aerial Vehicles)system.However,the development of the existing semantic SLAM is not mature,and its implementation is too dependent on the deep learning theory that requires high computing power,so it is not suitable for multi-rotor UAVs with certain load restrictions.Therefore,the thesis design a RGBD-based UAV semantic SLAM prototype system to complete the UAV trajectory estimation and construct a three-dimensional semantic map of the environment by improving the RGBD-SLAM with high accuracy tracking accuracy with adopting a light-weight target detection module and a three-dimensional semantic map construction method.Firstly,an RGBD camera is selected as the only visual sensor of the drone semantic SLAM system,after analyzing the realization principle,advantages and disadvantages of classic visual SLAM solutions,taking the drone application platform into consideration,on the basis of ORB_SLAM framework,a system solution that fusing efficient target detection algorithm and 3D target Segmentation method to complete the UAV trajectory estimation and construct a three-dimensional semantic map is set.Secondly,in response to the problem that the current deep learning-based target detection algorithm requires high computing platforms,a lightweight target detection network based on YOLOv3-Tiny is designed.By adjusting the network structure and related hyper-parameters,it combines the public data set and specific experimental environment date set to train the network and obtains the weight file that meets the experimental requirements,and based on this,it completes the semantic perception of the environmental goals and lays the foundation for the semantic SLAM system.Thirdly,the current three-dimensional point cloud segmentation methods based on geometric constraints mainly use single-environment three-dimensional information and lack of object-level segmentation effects,an improved three-dimensional target segmentation method based on GrabCut and VCCS is proposed to make full use of color images anddepth information obtained by the RGBD camera.It realizes target segmentation of the environment,and build an accurate three-dimensional semantic map based on Octomap by combining the segmented target object with semantic information.Finally,the thesis builds a drone semantic SLAM prototypical system to verify and analyze the method designed in this thesis under the conditions of public data sets and specific experimental environment data sets.The results show that the UAV semantic SLAM prototype system designed in this thesis not only guarantees the accuracy of front-end visual SLAM positioning and tracking but also integrates environmental semantic information to build a more detailed 3D semantic map.In this way,the feasibility of the system is verified and a good foundation is set for the UAV to complete more advanced tasks.