摘要： 针对当前高分辨率的多路视频拼接系统速度慢、实时性能低的问题，提出一种基于CPU和GPU并行架构的多路高清视频拼接算法。该算法在传统基于方向的快速特征点检测和旋转不变的特征描述算法上进行改进，删除针对尺度不变性应用的图像金字塔模块，并使用基于重叠区的局部配准方法，将配准后的图像数据在GPU设备端进行并行融合。在GPU与CPU异步执行的原则上，实现CPU端当前帧图像的配准，与其前帧图像融合，且以并行方式执行。通过显卡端图像数据计算与图像渲染之间的共享缓冲区，完成帧图像的快速渲染。实验结果表明，在4路200万像素的网络相机环境下，该算法实现的全景拼接系统的视频帧率达到17 f/s，可满足大场景的实时性需求。
Abstract: A kind of parallel architecture based on CPU and GPU is proposed to solve the low speed in high-resolution of multi-channel video stitching system, and this design is implemented on the physical platform successfully. The image pyramid module is deleted and a partial overlap ROI is improved from the traditional Oriented Features from Accelerated Segment Test(FAST) and Rotated Brief(ORB) algorithm. The image data after registration on the CPU side are then fused on the GPU side parallelly. Since the GPU and the CPU device call the execution asynchronously, the ORB image registration of the current frame on the CPU side and the image fusion of the previous frame on GPU side are executed parallelly. At last, the image fusion data which is computed by Compute Unified Device Architecture(CUDA) can be shared with the render device of GPU. It is this optimization that enhances the speed of the render. The result shows that the frame of this panorama video-stitching system can achieve 17 f/s, which meets the real-time demand in large scene.
Oriented FAST and Rotated Brief(ORB) algorithm,
Compute Unified Device Architecture(CUDA)