Repeatability Evaluation of Keypoint Detection Techniques in Tracking Underwater Video Frames

Abstract

Keypoint(s) or corners, as a stable feature possessing the defined characteristics of a robust point of interest remain an active research field for machine vision researchers due to its applications in motion capture, image matching, tracking, image registration, 3D reconstruction, and object recognition. There exist different techniques for keypoint detection; this paper focuses on direct computation on the gray-level analysis of interest point detection because of its straightforward implementation. In this contribution, an objective comparison of 12 state-of-the-art keypoint detection techniques; an application to feature matching have been executed in the context of underwater video sequences. These videos contain noise and all geometric and/or photometric transformations. Experiments are led on 5 videos containing in total 10 000 frames, evaluating the repeatability of the keypoints detectors. These detectors are evaluated on these complex videos by computing statistics-based repeatability, but also as a function of the Zero-Mean Normalized Cross-Correlation (ZNCC) scores.