| Positioning with and without control points are both important and irreplaceable in the domain of remote sensing and surveying and mapping.Ground control points with high positioning accuracy play an important role in improving the positioning accuracy of satellite images.Currently,most control points are obtained by manual deployment(calibration field)or feature extraction.The control points obtained by manual deployment are fixed in certain areas and have a high deployment cost,which require regular maintenance and on-the-spot GPS measurement.In the application,manual selection of points on the image to be processed is required,which needs a large amount of work.Feature extraction mostly uses artificially designed features.This method is limited by the quality of the reference image and has poor flexibility.In the application,it is necessary to match the reference image with the image to be processed.However,the difference in image sources,resolutions and imaging environment bring difficulties in image matching,which is not conducive to the generation and usage of ground control points on a global scale.According to the characteristics and application requirements of ground control,this paper focuses on the construction and application methods of global intelligent landmark control network.Global intelligent landmark control network refers to the ability of obtaining the classic surface features from ubiquitous geo-spatial information by intelligent methods,to extend or replace the traditional control points with the landmark control points set,and providing global three dimensional landmark control network with knowledge reserve and intelligence service for multi-scource remote sensing platforms such as space-borne and airborne(including unmanned vehicles).The main purpose of building landmark control network is to generate landmark control points by using image data with high positioning accuracy or three-dimensional scene data as reference data to provide enhanced ground control for ground observation data with low positioning accuracy,and to provide spatial datum and the control basis of globally positioning for various kinds of remote sensing satellites that lose the support of GNSS.In order to improve the accuracy,reliability and intelligence level of global remote sensing positioning,information resource support is built through the iterative update and continuous improvement of the global intelligent landmark control network.The main innovations of this paper are as follows:1.Aiming at the unified organization and management of global landmark control point data,a global three-dimensional grid partition and coding system was constructed.A multi-scale integer encoding method suitable for the organization of land mark data is proposed,and the management strategy of landmark data is designed to realize the effective organization of large-scale landmark control points.The multi-scale integer encoding method proposed in the thesis can convert the three-dimensional data organization problems into one-dimensional indexing to solve it,which only uses integer addition,subtraction and bit-field operations to achieve fast retrieval of landmark data and greatly improve the data management efficiency.The paper uses this method combined with commercial spatial database to manage large-scale landmark data,and compares with threedimensional Geohash,three-dimensional R-tree and other methods.The experimental results show that the method proposed in this paper has obvious advantages in data import,index construction,and regional query.In the three experiments,the efficiency of the proposed method is respectively about 20 times,40 times and 20 times those of three-dimensional R-trees.In the regional query,the efficiency of the proposed method is approximately three to five times that of three-dimensional Geohash.In addition,with the increase of data complexity and data volume,the advantage of proposed method is more obvious.2.To effectively match the images obtained from small inclination or vertical view with the reference images,this paper proposes a two-dimensional landmark control points generation and application method based on images with high positioning accuracy.This method selects objects with wide distribution,obvious features and relatively fixed locations to detect as landmarks,and adopts artificial intelligence methods for the intelligent detection of landmarks on the images with high positioning accuracy.Then,using grids,the landmark distribution and landmark quality as constraints,the detected control points are filtered and classified to ensure the uniform distribution of landmark ground control points.Based on the automatic generation of landmark control points,a set of image matching methods suitable for landmark control points is designed to realize the application of landmark control points.The paper conducted experiments with road intersections as landmarks,constructed a road intersection dataset,improved the Center Net network,and compared with a large number of target detection methods.Finally,the application strategy of landmark control points is adopted to realize the application of landmark control points on different satellite images with low positioning accuracy.Experimental results show that the proposed method in this paper can realize the landmark detection with high accuracy and the detection accuracy is 96%(AP50),which can meet the requirements of the generation of landmark ground control points.Compared with traditional methods,this method converts the match between images to be processed and reference images into the match between landmark image blocks,which effectively narrows the match areas,reduces the match difficulties,and has better stability.The match between landmark image blocks can achieve higher sub-pixel accuracy and can meet the requirements of control information transfer.3.Because two-dimensional landmark control points cannot transfer the control information well for large deformation and inclination image,this paper proposes a three-dimensional landmark control points generation and application method based on structured three-dimensional scene data and corresponding two-dimensional images.To increase the diversity of the dataset,this method expands the dataset manually based on existing open source landmark image dataset.Based on the deep learning method,the intelligent detection of the three-dimensional landmarks on the twodimensional image is realized.On the basis of the detection,the conversion relationship between the two-dimensional image and the three-dimensional model is studied,which realizes the automatic extraction of the three-dimensional landmarks.Constrained by grids,the distribution and size of landmarks,and positioning accuracy,the automatic generation,screening and quality grade determination methods of three-dimensional landmark control points are proposed.Based on the automatic generation of three-dimensional landmark control points,the three-dimensional landmark model is projected again according to the imaging parameters from images to be processed to acquire images with higher similarity to the images to be processed,which realizes the information transfer by match between re-projected images and images to be processed.This paper selects buildings as landmarks,and designs the generation and application experiments of landmark control points.The experimental results show that this method can effectively generate the three-dimensional landmark ground control points and achieve sub-pixel matching accuracy.Compared with the traditional methods,the proposed method in this paper reduces the differences between images and effectively promotes the stability of control information transfer by the re-projection of buildings,which can provide high-precision ground control point data for large inclination ground observation data,and realize the improvement of the image positioning accuracy. |
PDF Full Text Request