Research On Key Technologies Of Vision-based Geo-targeting With Unmaned Aerial Vehicle

Precisely striking interested targets is a key feature of IT-based warfare,while precisely scouting and positioning the targets is the basic prerequisite of precise strike.In recent years,driven by the combat idea of "no casualties,non-contact war",precise striking based on unmanned aerial vehicle(UAV)platform has been rapidly developing with precise positioning as its core technology to promote a combat style revolution.On the other hand,with the increasing maturity of the precision-guided weapon,which is represented by the joint direct attack bomb(JDAM),precise target positioning becomes an important issue that affects the fighting capacity.This paper mainly discussed the scenario that UAV used the equipped visible light or infrared camera to precisely position the target on the ground based on the photogrammetry principle,and researched the related key technologies including camera calibration,multi-view bundle adjustment,and image matching.The main work and innovative points of this paper include:1.For the camera installation relation calibration problem,this paper studied the non-intervisible pose estimation technique with the assistance of plane mirror,and proposed two kinds of non-intervisible pose estimation algorithm based on improper rotating average: a closed form solution for minimizing the 2 norm of chord distance and an iterative method for minimizing the 1 norm of geometric distance.The solution of minimizing the 2 norm of chord distance is superior to the traditional closed form solution in both computational efficiency and computational accuracy.The solution of minimizing the 1 norm of geometric distance first provided the robust algorithm for non-intervisible pose estimation.2.Image matching is the core basic issue of computer vision and photogrammetry,and it is also the key technology to improve the ground positioning accuracy of multipoint multi-view bundle.This paper performed exploratory research on sequence image matching,and proposed an innovative image matching method based on deep convolutional neural network and unsupervised learning,totally different from the traditional method.Based on the analysis of the dual relationship between two traditional problems,image interframe interpolation and sequence image matching,this paper used the method of "composite first,analyze later" to match the sequence images.In the "composite first" part,the deep convolutional neural network was used to interpolate values between the two input frames to be matched,and then an interframe image was generated.In the "analyze later" part,the gradient back propagation algorithm was used to analyze the sensitivity of the neural network in the interframe image to obtain the image matching information.Experiments conducted based on public test sets,such as KITTI and MPI-Sintel verified the effectiveness of the proposed method.It is worth emphasizing that: since a large number of video data had timing consistency naturally,it could be used for the training of neural networks mentioned in this paper without manual labeling.In other words,the method proposed in this paper enabled the computer to obtain sequence image matching by only watching the video,showing broad prospects.3.For parameter description of bundle adjustment in photogrammetry,a bundle adjustment method based on plane induced parallax was proposed.Through the reasonable assumption that the reference frame image was extracted without error,the plane induced parallax was used to describe 3D coordinates.Under the same configuration condition,the number of optimization parameters was greatly reduced compared with the traditional bundle adjustment method.The advantage is more obvious when the number of points in the world is much greater than the number of images.4.By analyzing the main error sources of the traditional ground positioning method that was based on the line of sight intersection,this paper studied a method to improve the accuracy of the ground positioning based on the multi-point and multi view constraints.The method makes full use of the geometric constraint,which is constructed by the accurate camera position and multiple view forms provided by airborne navigation system,to overcome the navigation system pose measurement error of the traditional method.Then it is possible that high precision ground positioning could be archived with low cost hardware equipment.After the geometric solvability of the method was carefully analyzed,the series bundle adjustment algorithms that performed precise ground positioning using this method were deducted,including the camera pointing system error correction algorithm in the case that camera parameters were known,and the algorithm correcting intrinsic parameters,such as the camera focal length and principal point.The theoretical accuracy of the method had been analyzed by using the Cramer-Rao lower bound,and the validity of the method had been verified by numerical simulation experiment and hardware-in-the-loop simulation experiment.