一种基于GPU的快速半全局优化深度图计算方法

发布时间：2018-06-02 03:16

本文选题：维重建 + 平面扫描　；参考：《工程科学与技术》2017年06期

【摘要】：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。
[Abstract]:Because of the large scale of image set and abundant matching information, fast and accurate multi-view stereo matching is seriously restricted by computational efficiency. To solve this problem, a fast semi-global optimization depth map calculation method based on GPU is proposed. Firstly, the initial matching cost of single image is calculated by plane scanning method on CPU. Then, the GPU semi-global optimization parallel computing architecture is proposed to aggregate the matching costs. The core algorithms are as follows: the global aggregation of task flows in different directions parallel to improve the utilization of the core processor; In order to avoid bandwidth limitation, parallel processing is realized by assigning each pixel computing task to each thread block accurately, and the data reuse on GPU is emphasized. Then GPU filter is used to eliminate the mutation point for image enhancement. Finally, the consistency of 3D spatial points on each depth image is taken as the constraint of outliers detection and optimization. The test results on multiple data sets show that the maximum computing speed of this method is 22.41 times that of the two-thread implementation method in the multi-core CPU system and 9.13 times of that of the 8-thread open implementation method, and the accuracy of the method is similar to that of the two methods. The results of comparison with similar depth map calculation methods show that the acceleration effect of this method is 5 times or more than that of other algorithms in the reconstruction process, and the open source point cloud comparison software is used to compare with other algorithms on the standard test data set. It is verified that this method can effectively improve the accuracy and integrity of the reconstruction results.
【作者单位】：四川大学计算机学院;
【基金】：国家自然科学基金资助项目(61571313) 四川省科技创新苗子工程资助项目(2016017)
【分类号】：TP391.41

【相似文献】