基于自适应投影矩阵的实时视频拼接算法

A REAL-TIME VIDEO STITCHING ALGORITHM BASED ON ADAPTIVE PROJECTION MATRIX

提出一种快速的算法用于高质量的实时拼接视频。算法分为两个阶段:后台阶段和实时阶段。后台阶段每隔一定周期运行,从同步的视频流中提取图像特征,采用RANSAC和LM算法计算投影矩阵。在实时阶段,采用从前一帧中得到的特征点点对,用光流法跟踪匹配点对的方法,对跟踪得到的点对由投影矩阵计算得到图像误差。若误差超过一定阈值或跟踪到点对数目太少,后台阶段就会再次执行。一旦得到了投影矩阵,就采取一种非线性的融合算法对视频进行融合。通过以上步骤,即使摄像头移动,算法也能运行快速。实验结果显示该算法大大改善了速度,而且拼接质量也很好。

In this paper,a fast algorithm is proposed to stitch video in real-time with high visual quality.This algorithm can be divided into two stages： the background stage and the real-time stage.The background stage runs every few seconds.In this stage,the algorithm extracts frame features from time-synchronised video streams and estimates projection matrix using RANSAC and LM algorithm.In real-time stage,the accumulated error in regard to the tracked pairs of key points is calculated by projection matrix,where the pairs of key points are tracked using optical flow algorithm and the pairs of feature points used are got from previous frame.If the error is larger than certain threshold or there are not enough pairs of key points tracked,the background stage will be run once again.Once the projection matrix is got,an effect nonlinear blending method is taken to blend the videos.By using these procedures,this algorithm can work fast and well even the cameras are moving.Results of experiment show that this algorithm can improve stitching speed greatly with good quality.