Research On Key Technologies Of UAV 3D Map Construction Based On Semantic Information In Dynamic Environment

With the continuous development of UAV technology,people have begun to apply UAVs in all walks of life,especially in the current situation of the raging new crown epidemic,UAVs have played an important role in epidemic prevention and control.However,whether it is for epidemic prevention and sterilization or material distribution,UAVs are required to have autonomous perception capabilities,that is,to know their own location and to correctly perceive the surrounding things.An accurate three-dimensional map is an indispensable condition for autonomous perception of UAVs.SLAM(Simultaneous Localization and Mapping)as the core technology of intelligent body represented by unmanned aerial vehicles and unmanned vehicles,can solve the problem of accurate 3D map construction to a certain extent,especially visual SLAM,which only needs the camera as the core technology.The advantages of sensors have a wide range of applications on UAVs.However,the visual SLAM system applied to UAVs at this stage is not robust to the outdoor dynamic environment.Dynamic instances in the dynamic environment will seriously affect the pose estimation of the camera,making the constructed map inaccurate and unable to be applied downstream.tasks(navigation and obstacle avoidance,etc.).Therefore,based on the ORB-SLAM2 framework,this thesis designs a set of UAV SLAM system RO-SLAM with better robustness in outdoor dynamic environment.Firstly,the classical semantic segmentation network PSPNet is improved for the problem of dynamic instance feature point removal,which reduces the amount of its computational parameters and facilitates deployment on mobile platforms such as UAVs.A new method based on feature enhancement and weighted fusion is proposed.The small target branch of,improves the detection accuracy of small targets in outdoor scenes.Semantic segmentation network is used to identify movable instances in the scene,remove their feature points,and do not participate in the pose estimation of the camera,so as to achieve the purpose of resisting dynamic interference.Since many holes will be generated in the constructed map after dynamic instance removal,the traditional method is to fill the holes through the multi-view geometric filling method,but this method is greatly affected by the pose estimation,and when there is no current The filling effect of the feature points of the frame will also be greatly reduced.Aiming at the above shortcomings,this thesis proposes a filling method based on generative adversarial network,which uses auxiliary descriptors to fill in static scenes according to semantic information,which enhances the fault tolerance rate of SLAM system.Finally,this thesis validates the semantic segmentation network and SLAM system on public datasets,both of which achieve good performance on public datasets.At the same time,this thesis also builds a verification platform to verify the system in a real scene.The verification results show that the system designed in this thesis can complete the construction of 3D maps in an outdoor dynamic environment.