Experimental analysis in visual features generation for submarine images preprocessed with contrast enhancement techniques

Abstract

[eng] Autonomous navigation in underwater environments performed by an AUV is a crucial activity in the context of the ARSEA project as it seems to be the best approach for mapping P.O. colonies. The navigation is performed with a SLAM implementation that relies partially on Visual Odometry and Loop Close Detection, which are both based entirely on visual information registered with cameras. However, the inherent conditions of the P.O. environment difficult the extraction of visual features so it is crucial to find the best feature extraction algorithm if a precise navigation is desired. This project focuses in the study of different possibilities for feature detection and tracking and also different image contrast enhancement algorithms thatmay improve the overall navigation performance. Given a dataset of raw images and another four datasets with the enhanced images, a software is used to run the feature tracking process given a detector and a descriptor. The resulting data is then analyzed in order to find the best detector and descriptor algorithms as well as image contrast enhancement algorithm. The loop close detection process shares the same limitations as those in Visual Odometry as it also relies on visual information. Thus, a similar methodology is used to find the best algorithms with small differences because it uses feature matching which has more general visual constraints than feature tracking. This difference can lead to different results so an agreement may be needed at the end.