| As a common hot research direction in the fields of photogrammetry and computer vision,3D reconstruction technology mainly utilizes multi view 2D images to obtain geographic information in 3D space.The most crucial step is to determine the correspondence between images,and dense matching is one of the main research methods.In the context of the rapid development of digital image sensors,the improvement of image resolution has also brought certain difficulties to current dense matching.However,semi-global dense matching has become the main method for obtaining dense point clouds in the fields of photogrammetry and computer vision due to its advantages of high accuracy,high efficiency,and strong applicability.However,this algorithm still faces the problem of matching costs that can not reflect real differences,and fixed penalty parameters that can not take into account different terrain conditions.In response to the above issues,this article conducted the following research within the framework of semi global dense matching algorithms:(1)We propose an improved GRA-Census algorithm that integrates gradients.By comparing and analyzing the matching effects of different cost calculation methods in highresolution close-range and aerial images,this provides guidance for subsequent research and use of semi global dense matching algorithms.Then,in response to the shortcomings of the Census algorithm,an improved GRA-Census algorithm with gradient fusion is proposed,which solves the problem of Census algorithm being easily affected by central pixel interference while maintaining the edge features of the object.Experimental results have shown that this algorithm can achieve more robust cost calculation results.(2)A semi global dense matching algorithm considering image segmentation information is proposed.By analyzing the impact of penalty parameters in the cost aggregation step on the semi global dense matching effect,a multi-scale image segmentation algorithm is adopted to obtain complete image segmentation edge information,and then the penalty parameters are adaptively determined during the cost aggregation process: for boundary pixels,the penalty parameter is reduced to allow for significant disparity changes,achieving the goal of maintaining edge features;Maintain a large penalty parameter for non boundary pixels,suppress their disparity changes,and achieve the goal of smoothing disparity.The experiment shows that the algorithm in this paper ensures the completeness of point clouds in flat areas while making the features of disparity mutation areas more complete and the edges sharper.(3)Implement a parallel processing strategy that combines software and hardware to optimize matching efficiency.At the algorithmic level,an image pyramid matching strategy is used to constrain the pre-set disparity range in semi global dense matching algorithms,which improves search speed and overcomes the drawbacks of violent search.Then,at the hardware technology level,Open MP is used for multithreading parallelization processing,and SIMD is used for data parallel optimization.Finally,through experimental testing,it can be concluded that the optimized algorithm has a matching speed improvement of 10-40 times compared to the unoptimized algorithm.(4)Validate and analyze the algorithm proposed in this article using various types of highresolution image data.The effectiveness and reliability of the GRA-Census algorithm were verified by comparing and analyzing the matching cost calculation methods from the ground close-up image dataset Middlebury to the large-scale aerial image dataset Vaihingen for urban scenes using these two types of standard datasets.Then,comparative experiments were conducted with the original SGM and SURE software in Vaihingen images and measured large array images,ultimately proving the superiority of our algorithm in standard datasets and measured data. |
PDF Full Text Request