Stability of Feature Detection Algorithms in Low Quality Video Encoding

Feature detection and description are the two enabling factors for image-based Augmented Reality, as they allow training on the images to be recognized for the superimposition of virtual objects. The aim of this paper is to evaluate the impact of codec and encoding parameters, which determine the quality of a video, on the stability and robustness of some of the most widely used methods for feature detection. By means of a frame-by-frame comparison between uncompressed YUV videos and the H.264/H.265 compressed versions, the stability of keypoints detected by ORB (Oriented FAST and Rotated Binary Robust Independent Elementary Features), SURF (Speeded-Up Robust Features), SIFT (Scale-Invariant Feature Transform) and BRISK (Binary Robust Invariant Scalable Keypoints) and the matches between them in the corresponding frames were evaluated. The results revealed the highest number of keypoint matches for SURF and ORB and the lowest number of matches for SIFT.