Research On Multi-Dimensional Measuring Techniques Of Unmanned Reconnaissance Aircraft For Ground Target

Unmanned reconnaissance aircrafts,as an important supplement to reconnaissance satellites and manned reconnaissance aircrafts,have gradually become an important strategy to comprehensively take control of the entire battlefield situation.Information warfare has become the main form of future wars,whose success mainly depends on its ability of information detection.The traditional way of spatial information presentation can no longer meet the needs of modern military reconnaissance,and it is necessary to obtain geospatial information by means of ground targets reconnaissance in multidimensional digitization.The purpose of this thesis is to use UAVs to obtain multi-dimensional reconnaissance information,conduct in-depth research on target positioning,speed measurement,semantic segmentation and modeling,and finally provide a theoretical basis for practical designs and operations.The main work and contributions of this article are as follows:This thesis analyzes the system structure and working mechanism of two types of measurement equipments,namely,Electro-optical Platform(EO)and Light Laser Detection and Ranging(LiDAR).The coordinate systems involved in the target measurement process based on the EO are defined,and the target localization equation is derived through the conversion relationship between the coordinate systems.The data structure of the point cloud is analyzed,and a semantic segmentation model based on Conditional Random Field(CRF)is established,and the method theory of 3D modeling using point cloud data is provided.For the deduced target localization equation,several error sources existing in the process of target localization are summarized and analyzed.Based on the Monte-Carlo method,the target localization error analysis model is established,and error factors are quantitatively analyzed one by one.The influence of each error factor on the target localization accuracy is then explored.For the systematic errors of the EO,a correction method of optical axis pointing error is proposed and the corresponding correction model is derived.The parameters in the proposed model is identified to reduce the non-linear errors and further increase the accuracy of groud target localization.On the basis of UAV target localization,Taylor equation and full differential ideas are utilized to derive the real-time velocity equation and error equation based on the change rate of the UAV attitude angle,camera optical axis pointing angle and other parameters.For the problem of maneuvering target state estimation,a dynamic space model is introduced into particle filter(PF),and a rejection method that uses non-cooperative game theory to eliminate negative particles in the process of particle filtering is introduced.The income matrix is defined according to the game concept,and the recommended density distribution function of the particle swarm is established to obtain the effective particle weight value.Then,the resampling of subsequent particles is guided to generate the correct sampling distribution.In the end,the relationship between the particle swarm and the target point reaches Nash equilibrium,and the target state is accurately estimated by the retained effective particles.This method can effectively reduce the degradation of particles and maintain the diversity of particle sets in order to reliably estimate the target states.To solve the problem of semantic segmentation with airborne LiDAR point cloud data,a semantic segmentation method is proposed that combines the location,contextual information and semantic information of point cloud data into a high-order CRF model.First,the optimal neighborhood selection algorithm based on entropy is used to obtain the optimal neighborhood of each point to obtain the statistical geometric characteristics of each point,and the Random Forest(RF)classifier is used to perform soft labeling with airborne LiDAR data.Based on the spatial context information between the point pairs,the soft labeling results are smoothed and corrected with the established weighted Potts model.Finally,during the construction of higher-order potential functions,a method for clustering nonground points is proposed,and based on the clustering blocks obtained by this method,the relevant semantic information of categories to be segmented is fused to establish a Sigmoid function based segmentation model,whose result is integrated into the higher-order CRF model as a higher-order potential function.By introducing the location,context information and semantic information of point cloud data into high-level CRF,it shows good semantic segmentation results,and provides a solution to the problem of category measurement of ground targets for unmanned reconnaissance aircraft.For three-dimensional modeling of building targets in airborne LiDAR data,a reconstruction method of building targets is proposed with the semantic segmentation results of airborne LiDAR data.Firstly,the building target points are extracted based on the semantic segmentation results,to obtain more reliable point cloud data after pre-processing.Then,the Alpha-Shape algorithm is utilized to extract the initial contour line of the building target point cloud,and a structured regular framework is established by proposing an edge line optimization method based on Gaussian Mixture Model(GMM),to obtain regular roof edge line.Finally,a plane slice extraction method based on regional growth is proposed,and accurate reconstruction of the building target with airborne LiDAR point cloud data is achieved by plane slice extraction and topological decomposition of the building target roof.