Comparing region-based and feature-based methods for ROV vision-based motion estimation

In this work, a comparison between different region-based and feature-based techniques used to estimate the motion of an underwater Remotely Operated Vehicle (ROV) is performed. In what respects region-based detectors, the article compares a previously analyzed template correlation technique with Maximally stable extremal regions (MSER). In previous works, the template correlation method proved to be the best (both in robustness to noise and computational time) when compared with several feature detectors (and descriptors) namely Scale Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), Center Surround Extrema (CenSurE), Features from Accelerated Segment Test (FAST), Binary Robust Independent Elementary Features (BRIEF) and Oriented FAST and Rotated BRIEF (ORB). Therefore, the need of comparing it with other region-based detectors arises. Nonetheless, previously untested detectors are now tested combined with BRIEF descriptors due to the good results obtained with BRIEF descriptors in previous works.All the algorithms are implemented in the same free open source library in order to achieve a fair benchmarking, in particular in terms of computational cost. The same experimental data set tested previously is used now in order to allow a relative comparison between these approaches as well as with previous approaches.The qualitative results show that MSER is unsuitable for this application while quantitative results proved that a combination of BRIEF descriptors and a variation of the CenSurE detector is faster and as robust to noise as template correlation. This is an important result as many techniques have been tested so far and all were always slower than template correlation.