Semantic-based visual vocabulary for loop closure detection

摘要

Recognizing when a robot is navigating in a previously visited location, known as loop closure detection, constitutes an essential task within any simultaneous localization and mapping (SLAM) system. This ability permits the agent to reduce the accumulated drift and generate a consistent map when estimating its traversed path. As the semantics of a scene attract the researchers’ attention, mainly due to their offered possibilities in allowing an intelligent robot to understand better what is included in its explored area, several methods for extracting this information have emerged. Aiming to recognize a familiar place and address the task of loop closure detection using appearance data provided by a camera sensor, this work proposes a novel pipeline based on high-level visual semantics. By classifying a scene through a finite number of classes, an n-gram-based semantic visual vocabulary is created using the combined data of sequential images. By comparing the query, i.e., the most recent captured image, with the previously visited set, votes are distributed into the database, while the most prominent ones are selected for further validation. The final decision comes via a geometric verification between the chosen images. High performances are achieved as our pipeline is evaluated on different publicly-available and widely-used image-sequences, surpassing several state-of-the-art algorithms.