多视匹配方法的计算任务分析及其GPU并行实现  

Computing Task Analysis of Multi-view Matching Method and GPU Parallel Implementation

作　　者：张丽[1,2,3] 汤晓涛[1,2] 戴晨光 纪松 胡国军[1,2,3] 

机构地区：[1]信息工程大学地理空间信息学院,河南郑州450052 [2]西安测绘研究所,陕西西安710054 [3]地理信息工程国家重点实验室,陕西西安710054

出　　处：《测绘科学技术学报》2013年第5期480-483,488,共5页Journal of Geomatics Science and Technology

基　　金："十二五"预先研究项目(40601030201);中国测绘科学研究院国家测绘地理信息局重点实验室开放基金项目(201202)

摘　　要：构建了多视匹配过程的总计算量模型,根据模拟参数赋值结果,分析得到了其中的密集计算任务,探讨了其GPU并行加速的必要性;针对单立体影像匹配技术细粒度GPU并行计算方案的不足,研究并设计了一种多视匹配密集计算任务的GPU粗粒度并行计算方案;利用专业级的GPU并行计算平台,对GPU粗粒度并行计算方案进行了实验验证,结果表明,该方案对于多视匹配过程中密集计算任务的并行加速效果十分显著。A total computing amount model was constructed for multi-view matching process.According to simulation parameter assignment results,the dense computing tasks exiseted in multi-view matching process were analyzed and obtained,and the parallel computing necessity was demonstrated.Since the fine grained GPU parallel computing scheme of single stereo matching technique was insufficient to get a high acceleration rate,a coarse grained GPU parallel computing scheme was studied and then designed for the dense computing tasks.Using professional GPU parallel computing platform,the coarse GPU paralle computing scheme was verified by experiments.And the experimental results proved that,for the dense computing tasks exsited in multi-view matching process,the GPU acceleration effect was quite significant.

关 键 词：多视匹配 计算任务 图像处理单元 并行计算 加速比 

分 类 号：P237[天文地球—摄影测量与遥感]