This paper proposes an evaluation of the impact of the 2D feature correspondence stage on an underwater vision-based navigation solution using a monocular Visual Odometry (VO) algorithm for linear velocity estimation. In particular, this work compares three different mismatch removal methods: the Cross-Check (CC), the Lowe’s Ratio Test (RT), and the Grid-based Motion Statistics (GMS). The performance of the three methods was assessed using two datasets containing real underwater images, which were collected by an Autonomous Underwater Vehicle (AUV) during monitoring activities over two distinct marine areas exhibiting different seafloor characteristics. The comparison is conducted considering both quantity and quality of features returned by the three approaches. In addition, the influence they have on the overall VO algorithm in terms of linear velocity accuracy is taken into account, using doppler velocity log readings as a reference. The results show that the three techniques are comparable in the case of a seabed characterised by identifiable and discernible features. In contrast, when surveying a more challenging and variable scenario, the RT technique shows a greater ability than CC and GMS to filter out erroneous 2D correspondences. This ensures higher accuracy in estimating the AUV linear velocity by the monocular VO algorithm. Furthermore, in both scenarios analysed, the RT technique is also the one that leads to a lower computational cost of the entire VO algorithm, and thus a better suitability for a real-time application onboard the AUV.

"File: https://ieeexplore.ieee.org/abstract/document/10244520"