一种基于GPU的快速半全局优化深度图计算方法

The accuracy and efficiency of multi-view stereo matching are seriously hindered by the scale of the image datasets and the wide variety of matching information.To address this issue,in this paper,a fast semiglobal optimizing method based on GPU was proposed for depth map calculation with parallel computing technology applied.First,the matching cost of single image was computed by the plane sweep method on CPU. Second,the matching costs were aggregated by using semiglobal optimizing method executed on GPU in parallel.The core module included two strategies:1) by aggregating the cost in each direction globally and parallelly,the utilization rate of the multi-cores of the GPU was increased;2) The computation on each single pixel was assigned to different threads and computed in parallel.To alleviate the constraint of limited bandwidth,the data reuse of GPU had also been taken into consideration.Then,the filtering on GPU was proposed to remove the outliers and enhance the image.Finally,the consistency of the 3D point projections on every image was used as a constraint for outlier detection and optimization.The proposed method had been tested on multi-group datasets.Experiments showed that the proposed method achieved as high as 22.41 times and 9.13 times speedup compared to two-threaded and eight-threaded implementation on a multi-core CPU system respectively,while also preserving similar quality.Compared with other algorithms of depth map calculation on tested datasets,the proposed method always achieved as fast as more than 5 times speedup in the reconstruction process.Compared with other algorithms on several datasets via open source pointcloud comparison software (CloudCompare),the proposed method could effectively improve the accuracy and integrity of the reconstruction results.