Underwater place recognition in noisy stereo data using FAB-MAP with a multimodal vocabulary from 2D texture and 3D surface descriptors

One fundamental problem in Simultaneous Localization and Mapping is place recognition, used to create loop closures, necessary to improve map quality. Ideally, the sensor data alone should be used to detect loop closures, referred to as appearance-based methods. Such methods usually employ feature descriptors, especially popular in underwater research are 2D visual features, and the bag-of-words approach. Here we demonstrate the combined use of 2D visual and 3D surface features, which is specifically interesting in the context of 3D underwater mapping with stereo data. Different sensor and feature types excel at different situations, and thus would benefit from a joint treatment in a bag-of-words based place recognition method. Specifically, this paper presents a method to employ multiple dictionaries computed from different feature types, extracted from different raw data in a single FAB-MAP-based place recognition method. Experiments on an underwater dataset show an increase of performance of the method when using a combination of feature types instead of single feature types.