现今的特征点轨迹稳像算法都是基于网格变形达到稳定视频的最终目的, 而保证结果不扭曲失真且稳定的网格变形需要由一定数量的长特征轨迹通过相应最优算法来实现。目前所提出的算法无法在保证良好时间性能下达到这一要求, 针对这个问题, 提出一种基于特征点轨迹增长的视频稳像算法。首先提取特征轨迹, 为避免算法优先选择较长轨迹而导致轨迹分布过于集中造成局部抖动的问题出现, 将特征点位置分布与轨迹长度相结合作为选择策略使特征点轨迹分布更加均匀; 接着利用低秩矩阵迭代逼近原理生成虚拟轨迹来实现轨迹增长; 最后利用网格变形生成稳定帧。将本文的算法与另外两种典型的特征点轨迹稳像算法相比较, 其中包括基于对极几何点转移的稳像算法以及基于三焦点张量重投影的特征点轨迹稳像算法。实验结果表明, 本文算法的特征点分布均匀且轨迹利用率高, 与基于对极几何点转移的稳像算法相比, 稳像效果更稳定并且时间复杂度更低, 与基于三焦点张量重投影的特征点轨迹稳像算法相比, 在保证稳像效果的同时时间复杂度更低。

Currently, image stabilization algorithms using feature point trajectory are all achieving the ultimate stable video grid deformation, and the only way to conduct a stable grid deformation is to guarantee a considerable number of long feature trajectories and put them into a proper series of algorithms. However, the proposed algorithms cannot achieve this requirement when a lower time complexity is needed. To address this problem, an image stabilization algorithm using feature trajectory augmentation is proposed. First, this algorithm extracts feature point trajectories. Then, to prevent the algorithm from selecting longer trajectories primarily and then causing a concentrated locus distribution that may cause a local jitter problem, the algorithm combines the feature point locus distribution with the track length as a selection strategy to make the distribution of feature points more uniform. Additionally, the algorithm generates virtual trajectories using a low-rank matrix iterative approach to achieve trajectory growth. Finally, it generates stable frames using grid transformations. The algorithm is compared with two other typical feature point trajectory image stabilization algorithms: the feature trajectory stabilization algorithm using epipolar point transfer and feature trajectory stabilization algorithm using trifocal tensor reprojection. As experimental results show, the feature points of the proposed algorithm are evenly distributed, and the trajectory utilization is significantly high. Compared with the feature trajectory stabilization algorithm using epipolar point transfer, the stabilization effect of the proposed algorithm is more stable, and the time complexity is lower. Compared with the feature trajectory stabilization algorithm using trifocal tensor reprojection, the time complexity is lower while ensuring the same stability.