基于改进SURF和P-KLT算法的特征点实时跟踪方法研究

 针对视频序列中运动目标的实时跟踪问题,提出一种基于改进SURF算法和金字塔KLT算法相结合的特征点跟踪方法。首先人工标定目标区域,利用改进的SURF算法分块快速提取具有高鲁棒性、独特性的特征点;然后在后续帧中应用金字塔KLT匹配算法对特征点进行稳定跟踪,采用基于统计的方法剔除错误匹配对;最后利用Greedy Snake分割算法提取轮廓确定更加精准的位置信息,更新目标区域。为使算法更具鲁棒性,还设计了离散点筛选、自适应更新策略。利用飞行视频数据库进行了大量的仿真,结果表明:该算法适用于多尺度图像序列中位置、姿态发生快速变化且结构简单的飞行器的稳定跟踪。帧平均时间为31.8 ms,比SIFT+P-KLT跟踪算法减少47.1%;帧几何中心、目标轮廓面积平均误差分别为5.03像素、16.3%,分别比GFTT+P-KLT跟踪算法减少27.2%、56.9%,比SIFT跟踪算法减少38.6%、68.4%。

Research of a Real-time Feature Point Tracking Method Based on the Combination of Improved SURF and P-KLT Algorithm

In order to track moving targets of image sequences in real time, a novel feature point tracking algorithm is proposed by using a combination of improved speeded up robust features (SURF) algorithm and pyramid kanade-lucas-tomasi (P-KLT) matching algorithm. First, the target box is marked manually and improved SURF algorithm is applied to extract features which are robust and distinctive in different blocks. Then, the features are tracked stably by the hierarchically iterative matching of P-KLT algorithm and mismatched points are eliminated utilizing a statistical method. Finally, the exact location of the target is obtained with the application of extracting the target contour by the Greedy Snake algorithm and the target box is updated automatically. Furthermore, discrete feature filter and adaptive feature updating strategy are designed to improve the robustness. Simulation results show that the algorithm can adapt to objective changes in attitude and size and track stably aerial vehicles with simple structures. Time consumption per frame is 31.8 ms, which is 47.1% less than SIFT+P-KLT algorithm. Geometric center error per frame is 5.03 pixel, which is 27.2% less than GFTT+P-KLT tracking algorithm and 38.6% less than SIFT+P-KLT tracking algorithm. Contour area error per frame is 16.3%, which is 56.9% less than GFTT+P-KLT tracking algorithm and 68.4% less than SIFT+P-KLT tracking algorithm.