摘要
为了解决高清视频的畸变校正及显示的实时性问题,提出了一种CUDA架构下的并行加速方案。系统利用张正友标定方法获得摄像机的内部参数和畸变参数,并利用GPU的大规模并行计算能力加速校正过程。校正后,位于显存的图像数据直接利用OPENGL驱动进行显示。针对不同架构GPU片上资源限制不同,设计了一种并行划分参数自整定算法,保证了程序移植到不同GPU后能充分利用硬件资源,实现最佳性能。实验结果表明,本文设计的系统对传统串行处理系统的综合加速比最高可达39倍以上,对2 596×1 920分辨率视频下的处理帧率可达100F/s以上。
Unistortion of the video image is a necessary step for accurate visual measurement and three- dimensional reconstruction, but excessive time-consuming is a serious problem of traditional serial process method for high definition video. To real-time calibrate and display the distorted video, a heterogeneous system with CUDA computing architecture is proposed in this paper. Zhangls calibration method is employed to get the camerars intrinsic and distortion parameters,undistortion maps are computed and uploaded to the display memory,and GPU processor with massive parallel computing ability is also used to accelerate the calibration process. To improve the display process, the direct driver technology from CUDA to OPENGL is used to output the corrected image data from the device memory. While the resource constraint and computation ability vary with the architecture of the GPUs, a self-tuning algorithm for parallel partitioning parameters is proposed to make that the application can get the best performance on different hardwares. Hardware platforms with different resources and process abilities are built to compare the undistortion and display process. The results of the experiment show the proposed system can get the speedup more than 39 with resolution of 2596 ×1920, and the frame processing rate can be more than 100 F/s.
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2013年第5期982-989,共8页
Journal of Optoelectronics·Laser
关键词
畸变校正
高清视频
GPU加速
并行计算
参数自整定
OPENGL
distortion correction high definition video GPU acceleration parallel computing parameter self-tuning OPENGL
