Learning a Deep Metric: A Lightweight Relation Network for Loop Closure in Complex Industrial Scenarios. Issue 1 (1st January 2021)

• Record Type: Journal Article
• Authors: Sheng, Jin; Liang, Chen; Yu, Gao; Changqing, Shen; Rongchuan, Sun
• Other Names: Liu Datong guestEditor.; Zhang Bin guestEditor.; (Peter) Liu Jie guestEditor.; Boland David guestEditor.
• Is Part Of: Chinese journal of electronics. Volume 30:Issue 1(2021)
• Page Start: 45
• Page End: 54
• Publication Date: 2021-01-01
• Subjects: Loop closure detection -- Relation network -- Image sequences -- End to end; Electronics -- Periodicals; Electronics -- China -- Periodicals; Electronics; China; Periodicals; 621.38105
• DOI: 10.1049/cje.2020.11.005
• Languages: English
• ISSNs: 1022-4653

Abstract

Abstract : The similarity metric in Loop closure detection (LCD) is still considered in an old fashioned way, i.e . to pre‐define a fixed distance function, leading to a limited performance. This paper proposes a general framework named LRN‐LCD, i.e . a Lightweight relation network for LCD, which combines the feature extraction module and similarity metric module into a simple and lightweight network. The LRN‐LCD, an end‐to‐end framework, can learn a non‐linear deep similarity metric to detect loop closures from different scenes. Moreover, the LRN‐LCD supports image sequences as input to speed up the similarity metric in real‐time applications. Extensive experiments on several open datasets illustrate that LRN‐LCD is more robust to strong condition variations and viewpoint variations than the mainstream methods.