基于改进SIFT和互信息法的单色和彩色视频高精度配准  

High precision registration of monochrome and color video based on improved SIFT and mutual information method

作　　者：谢万鹏 刘欢[1] 吴银花 陈雷[2] 张岩 王文娜 胡佳 韩军[1] XIE Wan-peng;LIU Huan;WU Yin-hua;CHEN Lei;ZHANG Yan;WANG Wen-na;HU Jia;HAN Jun(School of Opto-electronical Engineering,Xi'an Technological University,Xi'an 710021,China;School of Computer Science&Engineering,Xi'an Technological University,Xi'an 710021,China)

机构地区：[1]西安工业大学光电工程学院,陕西西安710021 [2]西安工业大学计算机科学与工程学院,陕西西安710021

出　　处：《液晶与显示》2023年第12期1689-1697,共9页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.12103039);陕西省自然科学基础研究计划(No.2021JQ-640);陕西省教育厅重点科学研究计划(No.21JY016);陕西省重点研发计划(No.2019ZDLGY16-01)。

摘　　要：为了解决单色视频和彩色视频的全自动高精度配准问题,设计了针对视频配准的首帧及后续帧配准方法,在保证配准速度的同时提高了配准精度。对于首帧,首先在尺度不变特征变换(SIFT)基础上,基于双摄像头的特点进行了改进,进而对单色视频和彩色视频进行了粗配准,增加了粗配准矩阵的获取稳定性和精确性;随后根据粗配准矩阵系数,使用尺度金字塔模式对单色视频和彩色视频的配准矩阵进行了配准修正。对于视频中的后续帧,采用具有步长金字塔的平移修正方法,对首帧中得到的修正配准矩阵进行再利用,并通过平移修正方法弥补了由不确定时差波动和相机摆动带来的平移误差,最终提高了单色视频和彩色视频的配准效率和配准精确度。实验结果表明,对于640×480大小的单色视频和彩色视频,本文算法相比传统SIFT算法将首帧配准均方根误差减少了0.79%,首帧标记点误差由约为1像素缩小到不可察,且连续视频帧的平均配准时间为0.357 s,同时仍然保持标记点误差不可察,较好地满足了单色视频和彩色视频配准的全自动、精度高、速度快、鲁棒性强等要求。In order to solve the problem of fully automatic and high-precision alignment of monochrome and color videos,the first frame and subsequent frame alignment methods for video alignment are designed to improve the alignment accuracy while ensuring the speed of alignment.For the first frame,firstly,the scale invariant feature transform(SIFT)is improved based on the characteristics of dual cameras,and then the coarse alignment is performed for monochrome and color videos,which increases the stability and accuracy of the coarse alignment matrix.Subsequently,the alignment matrices of monochrome and color videos are modified according to the coarse alignment matrix coefficients using the scale pyramid model.For the subsequent frames in the video,a translation correction method with a step pyramid is used to reuse the corrected alignment matrix obtained in the first frame,and the translation correction method is used to compensate for the translation errors caused by uncertain time difference fluctuations and camera oscillations,which finally improves the alignment efficiency and alignment accuracy of monochrome and color videos.The experimental results show that for monochrome and color videos of 640×480 size,the algorithm in this paper reduces the root mean square error of the first frame alignment by 0.79%and the marker point error of the first frame from about 1 pixel to undetectable compared with the traditional SIFT algorithm,and the average alignment time of consecutive video frames is only 0.357 s,while still keeping the marker point error undetectable.The algorithm in this paper better meets the requirements of fully automatic,high accuracy,fast and robustness for monochrome and color video alignment.

关 键 词：计算机视觉 视频配准 时空相似性 高精度修正 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]