Two-Pass K Nearest Neighbor Search for Feature Tracking

In recent years, feature tracking has become one of the most important research topics in computer vision. Many efforts have been made to design excellent feature matching methods. For large-scale structure from motion, however, existing feature tracking methods still need to improve in aspects of speed and matching confidence. To defense the drawbacks, in this paper, we design a simple and efficient feature tracking method based on the standard
$k$
nearest neighbor search. First, the parallel scale-invariant feature transform (SIFT) is selected as the feature detector to locate keypoints. Second, the principal component analysis-based SIFT-descriptor extractor is used to compute robust descriptions for the selected keypoints, in which normalized operation is used for boosting the matching score. Third, the two-pass
$k$
nearest neighbor search (TP-KNN) is proposed to produce correspondences for image pairs, then leading a significant improvement in the number of matches. Moreover, a geometry-constraint approach is proposed to remove outliers from the initial matches for boosting the matching precision. Finally, we conduct experiment on several challenging benchmark datasets to assess the TP-KNN method against the state-of-the-art methods. Experimental results indicate that the TP-KNN has the best performance in both speed and accuracy.