Efficient decentralized visual place recognition from full-image descriptors

摘要

Visual multi-robot simultaneous localization and mapping (SLAM) is an effective way to provide state estimation to a group of robots that operate in an unstructured and GPS-denied environment. This is a problem that can be solved in a centralized way, but in some instances it can be desirable to solve it in a decentralized way. Decentralized visual place recognition, then, becomes a key component of a decentralized visual SLAM system. Achieving it by having all robots send queries to all other robots would use vast amounts of bandwidth, and diverse approaches have been explored by the robotics community to reduce that bandwidth. In previous work, we have proposed a decentralized version of bag-of-words place recognition, which, by using a distributed inverted index, is able to reduce bandwidth requirements by a factor of n, the robot count. In this short paper, we instead propose a decentralized visual place recognition method that is based on full-image descriptors. The method consists in clustering the full-image descriptor space into several clusters and assigning each cluster to one robot. As a result, place recognition can be achieved by sending each place query to only one robot. We evaluate the performance of our new method versus a centralized implementation using the Oxford Robotcar and KITTI datasets and explore an inherent trade-off between performance and load balancing.