视频拼接中最优自适应单应性矩阵求解算法

针对交通场景中对视频拼接速度要求较高的特点,提出了利用视频帧最优单应性矩阵进行实时拼接的双线程快速算法。首先,利用SURF(Speeded up robust features)算法提取图像特征点;其次,通过NN(Nearest neighbor)算法以及优化的RANSAC(Random sample con-sensus)算法进行特征点的匹配,并去除误匹配点对;再次,利用重叠区域的归一化协方差相关函数最大化得到视频前k帧配准效果最佳的单应性矩阵,作为后继视频帧场景拼接的映射矩阵;同时,采用KLT算法对k+1的后继视频帧特征点进行动态跟踪,若匹配点对的数量变化超过了给定的阈值,则认为当前的最优单应性矩阵需要进行优化和变换,重新计算k幅视频帧中新的特征点对,经匹配后求取新的最优单应性矩阵。在交通场景中的拼接实验证明,该快速算法平均视频帧的拼接处理速度小于100ms,对存在旋转、尺度缩放、视角以及光照变化的图像都具有良好的拼接效果,具有参数估算准确,计算量小、速度快的优点,能够满足系统对视频拼接的实时性和精确性要求。

Algorithm for optimal homography matrix in video mosaic

We propose an algorithm for optimal homography matrix in video mosaic using double thread.First,the characteristic points are extracted using Speeded Up Robust Feature(SURF) method.Second,the corresponding characteristic points are matched using Nearest Neighbor(NN) and Random Sample Consensus(RANSAC) methods.Third,the optimal homography matrix of first k video frames is calculated using the normal covariance correlation of overlapped area between corresponding frames.The optimal homography matrix is taken as the mapping matrix of the succeeding frames.Meanwhile,the KLT algorithm is used for feature dynamic tracking of the frames succeeding k+1 frame.If the change of amount of the tracked feature points exceeds the given threshold,the current optimal homgoraphy matrix needs re-optimization and transformation;the pairs of feature points of k video frames are recalculated and matched to obtain the new optimal homgoraphy matrix.Image matching tests in traffic surveillance demonstrate that the average matching speed of the proposed algorithm is less than 100 ms;good matching quality is achieved for images with variable rotation,scaling,visual angle and illumination;parameter estimation is accurate and high speed with less computation,which meets the requirements of real-time and high accurate matching of the system.