Research On UAV SLAM Technology Based On Monocular Vision

With the continuous expansion of UAV(Unmanned Aerial Vehicle)application scenarios,the operating environment is becoming more and more complex and even unknown places.Therefore,the ability of intelligent perception of environment is becoming more and more important for intelligent UAV,which is also an important direction of the development of UAV.Simultaneous localization and mapping(SLAM),as the core technology of agent represented by UAV and unmanned vehicle,plays an important role in the field of intelligence.SLAM technology is the basis of intelligent perception and can solve the problem of robot positioning and environmental map rebuilding.Visual SLAM,especially monocular SLAM,is widely used on smart UAVs with more load constraints because it can use airborne cameras without additional sensors,provide rich and visible information,and can also be used to build maps and achieve location.However,the existed SLAM methods for monocular vision have some problems,such as too much map noise due to the influence of illumination,and inaccurate trajectory estimation,which have affected the application of this technology in unmanned aerial vehicles to a certain extent.Therefore,this paper first proposes a DoN(Difference of Normal)based point cloud back-end optimization method to overcome the shortcomings of direct spatial DSO(Direct Sparse Odometry)which is susceptible to light conditions.The method first rasterizes the point cloud and calculates the DoN difference operator of different scales in the point cloud in each raster.And then the raster with the smallest difference operator is selected for DoN filtering,so as to filter out the low-frequency noise information and ensure that the map point cloud data is unchanged,the map target is more intuitive and the detail features are more obvious.Then,this paper proposes an improved FAST method for adding line segment feature points to the problem that the semi-direct method SVO(Semi-direct Monocular Visual Odometry)has insufficient trajectory location estimation.This method increases the detection of line features by setting a threshold value,adds the endpoints of the detected segments to the FAST feature point set,and improves the accuracy of matching between adjacent frames by increasing the feature points,thus improving the accuracy of UAV trajectory estimation.Finally,this paper builds an unmanned aerial vehicle monocular vision SLAM verification platform based on the above contents to test and validate the proposed methods.The results show that under the number of similar point clouds,the map features reconstructed based on the improved DSO algorithm proposed in this paper are more obvious,and the information entropy value is higher.The improved SVO trajectory estimation algorithm can effectively improve the trajectory estimation,and more accurate UAV position can be obtained.