Research On Dense Map Creation Based On UAV's Monocular Vision And IMU Fusion

UAVs are small in size and have high spatial freedom.They occupy an indispensable position in exploration,surveillance,search and rescue scenarios,and have a wide range of application values.High-precision dense maps are the prerequisite for UAVs to complete positioning,obstacle avoidance and navigation tasks,and have attracted the attention of academia and industry.Traditional UAV-based dense mapping methods rely on the depth of field information measured by depth cameras,such as RGB-D cameras,and the wide range of applications is restricted.To this end,this article focuses on the creation of dense 3D maps based on the fusion of monocular cameras and inertial navigation measurement units(IMU).The main contributions are as follows:Aiming at the problem of low robustness of monocular visual pose estimation,considering that the IMU sensor is small in size,high in measurement frequency,and independent of external environmental characteristics,it is naturally robust to highfrequency jitter.Therefore,this paper uses IMU to solve the problem of pure vision.The accuracy of the pose estimation method under high-speed motion is reduced.Experiments show that the visual front end of the method in this paper can still track steadily under fast motion,while the well-known pure visual ORB-SLAM2 method loses its position.Aiming at the key bottleneck problem that monocular vision cannot obtain depth information,this paper uses traditional monocular depth estimation technology to first select representative pixels through a quad-tree structure for cost calculation and dynamic confidence propagation.After the preliminary depth extraction is completed,combined from The pixel depth value extracted from the high-texture area and the depth interpolation method are further filled in the depth map,and finally a dense depth map is generated according to the time fusion probability model for the creation of a three-bit point cloud map and a three-dimensional raster map.This is a traditional depth estimation method combining multiple methods.Experiments show that the method in this paper can still obtain a certain number of depth values stably in areas with weak textures and inconspicuous textures.The main innovation of this paper is to combine the improved SGM semi-global algorithm with the quad-tree structure pixel selection method,and obtain a reliable depth map by extracting high-texture area pixels,depth interpolation,and fusion time probability model map,and then matching it with The robust pose estimation scheme fused with monocular and IMU creates a real-time 3D map creation system on the UAV platform.