基于 GPU 的目标识别算法的并行化研究

摘要

针对可变形部件模型算法（DPM ）的计算量大，无法完成实时检测等问题，通过GPU编程模型CUDA ，在Nvidia GPU 上实现了 HOG 算法和DPM 算法的并行化；采用OpenCL 编程模型实现了DPM 算法在集成显卡上的并行化。通过CPU 和GPU 的协同计算，保证目标识别效果的前提下，并行化的算法的执行效率相比于OpenCV 中的CPU 或GPU 实现有明显的提高；通过对目标识别算法的并行化，结合其他算法，使得这类复杂算法能够在一些需要实时监测的工程领域中得到应用。%Aiming at the disadvantage of large amount of calculation ,the parallel solution methods of deformable part model(DPM ) algorithm and Histogram of Oriented Gradient (HOG ) algorithm based on GPU were proposed base on GPU with CUDA ,and the parallel solution method of DPM algorithm was also proposed based on integrated graphics card with OpenCL .With the cooperative computation of GPU and CPU ,under the premise of ensuring the target recognition effect ,the execution efficiency of the parallel al‐gorithms was significantly improved compared with the GPU or CPU implementations in OpenCV . Through the parallel implementations of target detection algorithms ,and combination with other algo‐rithms ,the target recognition algorithms can be applied in some engineering fields that need to be moni‐tored in real time .