摘要： 基于视频图像特征点配准的目标跟踪算法无法兼顾精确性、实时性和鲁棒性，针对该问题，提出一种基于特征位置预测与邻域一致性约束的视频特征快速配准算法。以标注点与目标标记框为模板，通过ORB特征匹配与邻域一致性检验，获得帧间标注点集的对应关系，并计算点集间的尺度变换以确定当前目标框，利用多帧已知标注点位置信息与运动连续性进行多项式回归预测，得到标注点集的位置。在此基础上，对特征点进行局部搜索、提取和描述，根据邻域一致性约束，利用邻域内的支持特征点集实现标注点的稳健匹配。实验结果表明，该算法可对多姿态目标特征点进行配准，与GMS、ORB、SIFT和SURF算法相比，该算法的实时性、准确性和鲁棒性明显提高。

关键词: 位置预测, 多项式回归, 运动连续性, 邻域一致性, 视频特征跟踪

Abstract: Existing target tracking algorithms based on video features points registration cannot balance precision,real-time performance and robustness.To address the problem,this paper proposes a fast video feature registration algorithm based on feature location prediction and constraints of neighborhood consistency.The algorithm takes labeled points and target label box as templates,and uses ORB feature matching and neighborhood consistency checking to obtain relations between inter-frame labeled point sets.The scale transform between point sets are also calculated to determine the current target box.The location of labeled point sets is obtained using polynomial regression prediction based on the location of known labeled points of multiple frames and motion continuity.On this basis,local search,extraction and description are performed on feature points,and robust matching of labeled points is implemented using assistant feature point sets in neighborhood according to neighborhood consistency constraints.Experimental results show that the proposed algorithm can implement registration of feature points of a multi-pose target,and has an obviously improved real-time performance,accuracy and robustness compared with GMS,ORB,SIFT and SURF algorithms.

Key words: location prediction, polynomial regression, motion continuity, neighborhood consistency, video feature tracking

中图分类号: 

• TP391