基于嵌入式平台双目测量的CUDA优化

为提高双目测量算法运算的速度,提出一种包括减少搬移、增大并行度和异步工作的方法来进行优化。基于快速双边滤波算法(fast bilateral stereo matching),它的匹配精度接近于全局匹配算法。对于标准测试样本,实现的方法在NVDIA TX1开发板上计算得到视差图所需要的时间更短,相比原来的CPU计算方法,代价聚合的效率有80倍的提升。实时双目测量方法为在嵌入式平台获取高质量双目视觉深度信息提供了有效、可靠的途径。

CUDA optimization method of binocular measurement on embedded platform

To accelerate the speed of binocular measurement algorithm, an optimized method was proposed, including reducing movement, strengthening parallelism and working asynchronously. It was based on fast bilateral stereo matching, a local algorithm whose matching accuracy was compared to that of the global matching algorithm. For standard test sets, the proposed method was implemented on NVIDIA TX1, which cost less time to get the disparity map. Compared with the original CPU implementation, the efficiency of cost aggregation is improved by 80 times. The real-time binocular measurement method provides an effective and reliable way to obtain high-quality binocular vision depth information on embedded platforms.