GPU加速与L-ORB特征提取的全景视频实时拼接

发布时间：2018-02-28 19:46

  本文关键词： 全景视频 图像拼接 异构计算 嵌入式GPU ORB　出处：《计算机研究与发展》2017年06期 　论文类型：期刊论文

【摘要】：全景视频是在同一视点拍摄记录全方位场景的视频.随着虚拟现实(VR)技术和视频直播技术的发展,全景视频的采集设备受到广泛关注.然而制作全景视频要求CPU和GPU都具有很强的处理能力,传统的全景产品往往依赖于庞大的设备和后期处理,导致高功耗、低稳定性、没有实时性且不利于信息安全.为了解决这些问题,首先提出了L-ORB特征点提取算法,该算法优化了分割视频图像的特征检测区域以及简化ORB算法对尺度和旋转不变性的支持;然后利用局部敏感Hash(Multi-Probe LSH)算法对特征点进行匹配,用改进的样本一致性(progressive sample consensus,PROSAC)算法消除误匹配,得到帧图像拼接映射关系,并采用多频带融合算法消除视频间的接缝.此外,使用整合了ARM A57CPU和Maxwell GPU的Nvidia Jetson TX1异构嵌入式系统,利用其Teraflops的浮点计算能力和内建的视频采集、存储、无线传输模块,实现了多摄像头视频信息的实时全景拼接系统,有效地利用GPU指令的块、线程、流并行策略对图像拼接算法进行加速.实验结果表明,算法在图像拼接的特征提取、特征匹配等各个阶段均有很好的性能提升,其算法速度是传统ORB算法的11倍、传统SIFT算法的639倍;系统较传统的嵌入式系统性能提升了29倍,但其功耗低至10W.
[Abstract]:Panoramic video is a video recording omnidirectional scene from the same viewpoint. With the development of virtual reality (VR) technology and live video technology, However, CPU and GPU are required to have strong processing capability. Traditional panoramic products often rely on large equipment and post-processing, which leads to high power consumption and low stability. In order to solve these problems, L-Orb feature point extraction algorithm is proposed, which optimizes the feature detection region of segmented video image and simplifies the support of ORB algorithm for scale and rotation invariance. Then the feature points are matched by the locally sensitive Hash(Multi-Probe algorithm, and the mismatch is eliminated by using the improved sample consistency progressive sample consensus algorithm, and the frame image splicing mapping relationship is obtained, and the multi-band fusion algorithm is used to eliminate the seam between videos. Using the heterogeneous embedded system of Nvidia Jetson TX1, which integrates ARM A57 CPU and Maxwell GPU, the real-time panoramic mosaic system of multi-camera video information is realized by using the floating-point computing ability of Teraflops and the built-in video capture, storage and wireless transmission module. The parallel strategy of GPU instruction block, thread and stream is used to accelerate the image mosaic algorithm effectively. The experimental results show that the algorithm has a good performance improvement in every stage of image stitching, such as feature extraction and feature matching. The speed of the algorithm is 11 times that of the traditional ORB algorithm and 639 times that of the traditional SIFT algorithm. The performance of the system is 29 times higher than that of the traditional embedded system, but its power consumption is as low as 10W.
【作者单位】： 武汉理工大学计算机科学与技术学院;交通物联网技术湖北省重点实验室(武汉理工大学);佛罗里达大学电气与计算机工程系;
【基金】：国家自然科学基金项目(61303029)~~
【分类号】：TP391.41
，

本文编号：1548695