Exploring Bi-directional Loop Closures in An Urban Environment

Abstract

Loop closure detection is a task of recognizing whether a robot or vehicle has returned to a previously visited location during the traversal. This process helps to correct accumulated errors, known as drifts, in robot’s position. By doing so, it ensures accurate localization and mapping of the environment, which are key components of simultaneous localization and mapping (SLAM) systems. Accordingly, the loop closure systems must yield reliable results, even in challenging conditions, as errors can harm to overall functionality and safety of the autonomous systems in real-world applications. Most of the existing research on loop closures focuses on cases involving changes in appearance or minimal viewpoint variations, where the scenes are observed from very similar perspectives in the same direction. Only a handful of studies have investigated bi-directional loop closures, which refers to the ability of the system to recognize a scene even when observed from the opposite direction. This problem has been relatively underexplored in research until very recently. The limited research on bi-directional loop closures suggests a need for further investigation into the topic in order to enhance the performance of loop closure detection task in such setting.

This work aims to investigate bi-directional loop closures with a particular focus on urban environments due to their relevance in reflecting the challenges and dynamics of real-world autonomous driving systems, while also providing findings for forest and in/outdoor settings. In this regard, Oxford RobotCar dataset has been utilized for its representation of an urban setting, capturing driving scenarios such as traffic, cars and pedestrians. The utilized approach is based on two separate Siamese networks, each designed for place recognition and pose regression tasks. The networks are trained in end-to-end manner using data pairs specifically prepared to support learning bidirectionality. The results indicated that while the approach was effective for loop closure detection in non-urban datasets, namely FinnForest and PennCOSYVVIO, its performance was less satisfactory on urban dataset, and this can be attributed to the increased complexity and variability of urban environments.