一种基于GPU的快速半全局优化深度图计算方法

A Fast Semiglobal Optimization Algorithm for Computing Depth Map with GPU

由于图像集规模巨大、匹配信息丰富,快速精准多视图立体匹配受计算效率严重制约。针对该问题,提出一种基于GPU的快速半全局优化深度图计算方法。首先,在CPU上通过平面扫描方法计算单张图像初始匹配代价。然后,提出GPU半全局优化并行计算架构,对匹配代价进行聚合,其核心算法为:在全局进行各方向聚合任务流并行以提升众核处理器的利用率;在局部通过将各像素计算任务准确分配到各线程块内实现并行处理,且注重GPU上数据重用以避免带宽限制。再通过GPU滤波剔除突变点进行图像增强。最后,将3维空间点在各深度图像上的一致性作为异常值检测和优化的约束条件。在多组数据集上测试结果显示,该方法计算速度最高为多核CPU系统中开启2线程实现方法的22.41倍,为开启8线程实现方法的9.13倍,且与两者精度相当;与同类深度图计算方法比较结果表明,该方法在重建过程中加速效果均为其他算法的5倍及以上;通过使用开源点云比较软件在标准测试数据集上与其他算法比较,验证了该方法能有效提高重建结果的精度和完整度。

The accuracy and efficiency of multi-view stereo matching are seriously hindered by the scale of the image datasets and the wide vari- ety 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：l） 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 en- hance 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 sottware （CloudCompare）,the proposed method could effectively improve the accuracy and integrity of the reconstruction results.