| Textured models of urban environments are needed for many photogrammetry and computer vision applications.These applications include but are not limited to urban development planning,3D scene classification,low flying automatic vehicle obstacle avoidance,etc.Therefore,the relevant research on using remote sensing imaging data to achieve the extraction of large range of building target structural information has always been a research hotspot in related fields.The method of combining airborne lidar with aerial optical image is becoming more and more popular in urban environment modeling.The generation process of LPC(LDAR Point Clouds)provides little vertical surface data,and most airborne LDAR sensors do not provide color information of the scene.Detailed structural information and optical information are needed to generate a realistic three-dimensional model of the urban environment.High resolution oblique aerial image can provide optical and structural information for airborne lidar effectively and has high plane accuracy.Therefore,this paper takes the advantage of the two data sources as the means to jointly realize the target structure extraction as the purpose,and carries out the following research work from the aspects of aerial image enhancement,multi-view aerial image registration,and the combined two data sources’ complementary extraction of the target structure:(1)For the traditional wavelet enhancement algorithm applied to aerial image processing effect is not ideal.There are many problems,such as the loss of image details while enhancing the image,weakening the edge information of the target in the image and reducing the contrast of the image.In order to maintain good edge detail information,highlight sensitive image information,and better reconstruct enhanced image,a new aerial image enhancement algorithm based on multi-view wavelet transform fusion is proposed.In this algorithm,the salient image and the detail image of the original image are fused by the wavelet transform,the image fusion criteria are weighed,and the weighted average of the low frequency signal and the variance of the high frequency signal are maximized respectively to get the reconstructed image.Experimental results show that this method has a good effect on aerial image enhancement.(2)In view of traditional convolution neural network to view aerial image registration training,failed to make full use of between multiple views as the connection between the edge character,in order to extract the multi-layer image edge structure between the characteristics of the information,put forward a kind of based on edge character and convolutional neural network combined aerial image registration method,through the window gray weighted algorithm to extract image edge character figure,and the edge character figure as a convolution of the neural network input for training,during the testing phase,give a pair of new visual images,more training model can predict the image after the corresponding relationship between the space.Experimental results show that the algorithm realizes image alignment transformation and improves the accuracy of image registration.(3)In the fusion stage of aerial image and LIDAR point cloud image,three fusion models are firstly summarized in detail.The fusion algorithm based on camera calibration is used to map ground point cloud into aerial image according to camera imaging principle.Based on the fusion algorithm of image registration,the coarse registration parameters are calculated by using the salient features extracted from the depth diagram of 3D model derived from lidar and aerial image,and the ICP(Iterative Closest Point)algorithm is used to refine the registration parameters further.A fusion algorithm based on machine learning is proposed to extract 2D corner points from images and 3D lidar models without texture,and then the corresponding 3D corner points corresponding to orthogonal structure are used as features.Then,the correspondence between images and 3D models is realized by using Hough transform and generalized M-estimation sample consistency.Then,the fusion simulation experiments were carried out by comparing the fusion algorithms of image registration and machine learning with the experimental data,and the fusion accuracy of each algorithm was calculated. |
