Applying Image Matching to Video Analysis

Dealing with the volume of multimedia collected on a daily basis for intelligence gathering and digital forensics investigations requires significant manual analysis. A component of this problem is that a video may be reanalyzed that has already been analyzed. Identifying duplicate video sequences is difficult due to differences in videos of varying quality and size. This research uses a kd-tree structure to increase image matching speed. Keypoints are generated and added to a kd-tree of a large dimensionality (128 dimensions). All of the keypoints for the set of images are used to construct a global kd-tree, which allows nearest neighbor searches and speeds up image matching. The kd-tree performed matching of a 125 image set 1.6 times faster than Scale Invariant Feature Transform (SIFT). Images were matched in the same time as Speeded Up Robust Features (SURF). For a 298 image set, the kd-tree with RANSAC performed 5.5 times faster compared to SIFT and 2.42 times faster than SURF. Without RANSAC the kd-tree performed 6.4 times faster than SIFT and 2.8 times faster than SURF. The order images are compared to the same images of different qualities, did not produce significantly more matches when a higher quality image is compared to a lower quality one or vice versa. Size comparisons varied much more than the quality comparisons, suggesting size has a greater influence on matching than quality.