Research On Estimation Method Of Scene Depth Information Based On Single Camera And Multi-view

Scene depth estimation is a hot research content in computer vision.The commonly used multi-view scene depth estimation is mainly based on linear arrays to obtain parallax through stereo matching.The improvement of accuracy is mainly reflected in the optimization of the matching and aggregation process.It is difficult to achieve new breakthroughs.In actual research,the arrangement of cameras in the image acquisition process The method requires high requirements,image calibration cannot be performed,and image parameter information is difficult to obtain.In order to overcome the limitations of the calibration process and further improve the scene depth estimation results,this paper proposes a nonlinear array to collect multi-view images for depth estimation.Due to the view array arrangement The randomness of the image,the parameter information between images is unknown,and the stereo matching method for finding the depth will no longer be applicable.Aiming at the difficult problem of nonlinear image array depth estimation,this paper proposes the idea of using the siamese neural networks and the plane sweeping algorithm to solve the depth.Firstly,siamese neural networks is used to learn image similarity to realize the cost calculation between images and then the cost aggregation estimation is used to get initial parallax range of the image.Because the obtained disparity results lack texture information,in order to further improve the depth estimation accuracy,based on the estimated depth range at this time,plane sweeping algorithm is used to directly map the depth range to obtain the mapping structure of the image within a limited range.Then use the optimization function to calculate the final parallax result layer to complete the further estimation of the image.The combination of the above two methods improves the image depth results,but the depth estimation process is based on two views,and there is a problem of incomplete information due to the lack of view information in the matching process.Aiming at the problem of incomplete depth information,this paper proposes a method based on multi-view stitching and fusion to complement the scene structure.First,the depth map obtained above is registered to obtain the matching between the two depth maps,and the overlap between the images is calculated according to the matching results Regions,then merge the overlapping regions to achieve depth map fusion.The results show that the method in this paper effectively overcomes the shortcomings of the two depth estimation methods,improves the accuracy of depth estimation,and enriches the texture information of scene depth estimation.