Research On 3d Reconstruction Of Ground Environment Based On Uav Images

With the development of science and technology,information detection in highrisk environments has become a research hotspot for researchers at home and abroad.It has become an irresistible trend for intelligent robots to replace workers in high-risk environments.With the increasing urgency of this demand,researchers have put forward a number of solutions,among which the Simultaneous Localization and Mapping(SLAM)technology is the best technology to complete complex environment detection tasks at this stage.However,the embedded chips of small intelligent robots,such as portable UAVs and unmanned vehicles,have limited computing power,and the existing SLAM technology is too cumbersome to be suitable for real-time operation on embedded devices.Therefore,it is necessary to deploy the SLAM framework in the embedded devices with limited computing power and reconstruct the 3D environment with rich information in real time.This thesis studies the 3D reconstruction technology of ground environment based on UAV.The specific work of this thesis is as follows:Firstly,a quadtree based ORB feature point extraction and matching algorithm is proposed to solve a series of illumination problems,such as illumination variation and uneven illumination,in the running environment of equipment.In the feature point extraction stage,the traditional quadtree equalization is added to the image enhancement algorithm to overcome the influence of light on the feature point extraction effect;In the feature point matching stage,the directional blocking feature of the quadtree algorithm is combined with the matching direction between feature point pairs to eliminate some mismatched feature point pairs and reduce the number of external points in the RANSAC iteration process.Experiments show that the improved algorithm weakens the influence of light on ORB feature extraction,and solves the shortcomings of RANSAC algorithm,which consumes long time and has poor matching accuracy.Then,aiming at the low initialization accuracy of the front odometer in SLAM system,a front odometer calculation method based on line characteristics and probability distribution is proposed.In the monocular initialization,the line feature is used to select the appropriate transformation matrix;Then,the mismatched feature points are classified by probability distribution principle and transformation matrix;Finally,accurate monocular initialization data is obtained.In the odometer operation link,the absolute scale of the map is recovered using the positioning and mapping algorithm coupled with vision and inertial navigation.The experimental results show that the improved algorithm can select the correct transformation matrix and has a good positioning and mapping effect.Finally,aiming at the low efficiency of the back end loop detection algorithm of SLAM system,a multi data fusion loop detection algorithm is proposed.Firstly,the correlation filtering algorithm and IMU information are applied to the loopback system,and the loopback results are judged by the comprehensive indicators of various conditions;Then,the same feature points at the loopback position are fused using pose transformation;Finally,BA optimization algorithm is used to optimize all key frames and map points in the whole process.Experiments show that the loop detection algorithm can ensure the matching efficiency,and the number of image frames required for loop back has been greatly reduced;The SLAM system built in this thesis can still loop back in the harsh lighting environment,and the trajectory error of the reconstructed map is relatively small.