Comparison of feature detection techniques for AUV navigation along a trained route

Autonomous underwater vehicles (AUV)s traversing a path will incur positional error drift over time while submerged. We are developing a route following system which is based upon features extracted from the seabed using sidescan sonar collected in a training phase. Through matching of sonar images, this system navigates over a path without the need for a continual global position estimate. At the core of this system is the need to reliably extract features and match images derived from the sonar. At our disposal is an array of algorithms which implement the OpenCV common interface for feature extraction and matching. Using pre-collected sets of data we compare the performance of several of these algorithms in the context of matching sonar image tiles. Our results compare the performance of various feature types over two common sets of data. The feature types tested include SIFT[12], SURF[3], MSER[13], STAR[1], ORB[15], and BRIEF[4].