|
|
Improved SeqSLAM Using Hierarchical Navigable Small World Graphs |
ZHANG Mengzhen, WANG Qingzhi, LIU Qipeng
|
Institute of Complexity Science, Qingdao University, Qingdao 266071, China |
|
|
Abstract SeqSLAM is a widely used loop closure detection algorithm in mobile robot and autonomous vehicle field. It could recognize revisited places by comparing sequences of images even under dramatic changes of season, illumination, and weather. However, SeqSLAM is vulnerable to viewpoint changes. In addition, SeqSLAM compares sequences of images by brute force method, which prevents its real-time application to large-scale image datasets. To address these problems, we first represent each image by a kind of low dimensional description — vector of locally aggregated descriptors (VLAD) which is robust to viewpoint changes, and then replace the brute force method by an approximate nearest neighbor search algorithm — hierarchical navigable small world graphs (HNSW). Tests on publicly available datasets show that, the improved SeqSLAM with VLAD and HNSW could obtain much better detection results in the respect of precision-recall evaluation and the search time is reduced by orders of magnitude. We make code publicly available at https://github.com/qipengliuNTU/Efficient-SeqSLAM-with-HNSW.
|
Received: 13 October 2021
Published: 19 April 2023
|
|
|
|
|
No related articles found! |
|
|
|
|