Research On Key Technologies Of 3D Scene Construction Based On Multi-view UAV Images

As a technique to restore the real-world,three-dimensional reconstruction from unmanned aerial vehicle(UAV)images is widely used in many practical fields.Recently,with the rapid development of UAV technology,high-precision multi-view UAV sequence images have become convenient to access,making it possible to realize highprecision 3D reconstruction of large-scale and large-scale scenes.Motivated by these,this thesis focuses on the 3D reconstruction of typical scenes with UAV sequence images.The main contributions are as follows:(1)Retinex-GAG,an image enhancement algorithm based on guided filtering and adaptive gamma correction for multi-view UAV sequence images is proposed to alleviate radiation degradation in UAV images,such as brightness distortion,shadow shading,and messy colour.Based on Retinex Net,the Guided-Filtering algorithm is introduced to denoise the image while retaining the detailed information.Meanwhile,the Gamma correction method is applied to adaptively correct the brightness distortion caused by excessive enhancement.In addition,L1 regularization operation is utilized to prevent overfitting.The comparison experiments performed on the NPU Drone Map dataset and Epfl Quartier Nord dataset verify the effectiveness of the model design.Compared with the traditional image enhancement algorithms like SSR,MSR,MSRCR and deep-learning-based method Retinex Net,Retinex-GAG can achieve better enhancement.(2)Considering the vast redundant matching pairs in UAV sequence images,this thesis designed the Enhanced-ITC topology connecting the network to extract the key matching pairs.Firstly,the spatial relational constraint algorithm is utilized to screen the matched image pairs.Firstly,this thesis proposes a preliminary screening method for image matching pairs using a spatial relationship constraint algorithm.Based on the importance of the overlap area to represent matching pairs,an initial topological connection is established and simplified using the maximum spanning tree algorithm.Then,the topological connection is extended along the flight direction and vertical direction to obtain an enhanced topological connection.Experimental results on the NPU Drone Map dataset and Epfl Quartier Nord dataset demonstrate that the proposed method can efficiently handle redundant matching pairs,effectively reduce the network’s computational burden,and improve the accuracy and efficiency of 3D reconstruction.Furthermore,the generalization ability of Enhanced-ITC was validated on the test image sets in Colmap.(3)This thesis proposes a depth estimation network based on a sequence of UAV images,called DLC-RMVSNet.The network uses dynamic multiscale feature extraction,lightweight cost volume,content-aware cost volume aggregation module,and RNN-CNN recurrent convolutional hybrid network to improve the accuracy and efficiency of the network.Tested on the NPU Drone Map dataset and Epfl Quartier Nord dataset,the results show that DLC-RMVSNet can obtain good 3D reconstruction results in typical scenarios.Compared with MVSNet,it has advantages in reconstruction,completeness,number of 3D reconstruction points,and running time.In addition,the generalization capability of DLC-RMVSNet network was verified on the Pix4 D image test sets.