Ensemble of bayesian filters for loop closure detection

Mohammed Omar Salameh, Azizi Abdullah, Shahnorbanun Sahran

Research output: Contribution to journalArticle

Abstract

Visual Simultaneous Localization and Mapping (vSLAM) system is widely used by autonomous mobile robots. Most vSLAM systems use cameras to analyze surrounding environment and to build maps for autonomous navigation. For a robot to perform intelligent tasks, the built map should be accurate. Landmark features are crucial elements for mapping and path planning. In the vSLAM literature, loop closure detection is a very important process for enhancing the robustness of the vSLAM algorithms. The most widely used algorithms for loop closure detection use a single descriptor. However, the performance of the single descriptors appears to worsen as the map keeps growing. One possible solution to this problem is to use multiple descriptors and combine them as in Naive and linear combinations. These approaches, however, have weaknesses in recognizing the correct locations due to overfitting and highbias, which hinder the generalization performance. This paper proposes the usage of ensemble learning to combine the predictions of multiple Bayesian filter models which make more accurate prediction than individual models. The proposed approach is validated on three public datasets; namely, Lip6 Indoor, Lip6 Outdoor and City Centre. The results show that the proposed ensemble algorithm significantly outperforms the single approaches with a recall of 80%, 97% and 87%, with 100% precision on the three datasets, and outperforms the Naive approach and the existing loop closure detection algorithms.

Original languageEnglish
Pages (from-to)235-242
Number of pages8
JournalPertanika Journal of Science and Technology
Volume25
Issue numberS6
Publication statusPublished - 1 Jun 2017

Fingerprint

filter
robots
Bayes Theorem
prediction
cameras
multiple use
Motion planning
learning
planning
Mobile robots
Learning
navigation
Navigation
Cameras
detection
Robots
Datasets

Keywords

  • Appearance-based localization
  • Ensemble learning
  • Loop closure detection

ASJC Scopus subject areas

  • Computer Science(all)
  • Chemical Engineering(all)
  • Environmental Science(all)
  • Agricultural and Biological Sciences(all)

Cite this

Ensemble of bayesian filters for loop closure detection. / Salameh, Mohammed Omar; Abdullah, Azizi; Sahran, Shahnorbanun.

In: Pertanika Journal of Science and Technology, Vol. 25, No. S6, 01.06.2017, p. 235-242.

Research output: Contribution to journalArticle

@article{5ad32c1d99b048b88f0fada707597b47,
title = "Ensemble of bayesian filters for loop closure detection",
abstract = "Visual Simultaneous Localization and Mapping (vSLAM) system is widely used by autonomous mobile robots. Most vSLAM systems use cameras to analyze surrounding environment and to build maps for autonomous navigation. For a robot to perform intelligent tasks, the built map should be accurate. Landmark features are crucial elements for mapping and path planning. In the vSLAM literature, loop closure detection is a very important process for enhancing the robustness of the vSLAM algorithms. The most widely used algorithms for loop closure detection use a single descriptor. However, the performance of the single descriptors appears to worsen as the map keeps growing. One possible solution to this problem is to use multiple descriptors and combine them as in Naive and linear combinations. These approaches, however, have weaknesses in recognizing the correct locations due to overfitting and highbias, which hinder the generalization performance. This paper proposes the usage of ensemble learning to combine the predictions of multiple Bayesian filter models which make more accurate prediction than individual models. The proposed approach is validated on three public datasets; namely, Lip6 Indoor, Lip6 Outdoor and City Centre. The results show that the proposed ensemble algorithm significantly outperforms the single approaches with a recall of 80{\%}, 97{\%} and 87{\%}, with 100{\%} precision on the three datasets, and outperforms the Naive approach and the existing loop closure detection algorithms.",
keywords = "Appearance-based localization, Ensemble learning, Loop closure detection",
author = "Salameh, {Mohammed Omar} and Azizi Abdullah and Shahnorbanun Sahran",
year = "2017",
month = "6",
day = "1",
language = "English",
volume = "25",
pages = "235--242",
journal = "Pertanika Journal of Science and Technology",
issn = "0128-7680",
publisher = "Universiti Putra Malaysia",
number = "S6",

}

TY - JOUR

T1 - Ensemble of bayesian filters for loop closure detection

AU - Salameh, Mohammed Omar

AU - Abdullah, Azizi

AU - Sahran, Shahnorbanun

PY - 2017/6/1

Y1 - 2017/6/1

N2 - Visual Simultaneous Localization and Mapping (vSLAM) system is widely used by autonomous mobile robots. Most vSLAM systems use cameras to analyze surrounding environment and to build maps for autonomous navigation. For a robot to perform intelligent tasks, the built map should be accurate. Landmark features are crucial elements for mapping and path planning. In the vSLAM literature, loop closure detection is a very important process for enhancing the robustness of the vSLAM algorithms. The most widely used algorithms for loop closure detection use a single descriptor. However, the performance of the single descriptors appears to worsen as the map keeps growing. One possible solution to this problem is to use multiple descriptors and combine them as in Naive and linear combinations. These approaches, however, have weaknesses in recognizing the correct locations due to overfitting and highbias, which hinder the generalization performance. This paper proposes the usage of ensemble learning to combine the predictions of multiple Bayesian filter models which make more accurate prediction than individual models. The proposed approach is validated on three public datasets; namely, Lip6 Indoor, Lip6 Outdoor and City Centre. The results show that the proposed ensemble algorithm significantly outperforms the single approaches with a recall of 80%, 97% and 87%, with 100% precision on the three datasets, and outperforms the Naive approach and the existing loop closure detection algorithms.

AB - Visual Simultaneous Localization and Mapping (vSLAM) system is widely used by autonomous mobile robots. Most vSLAM systems use cameras to analyze surrounding environment and to build maps for autonomous navigation. For a robot to perform intelligent tasks, the built map should be accurate. Landmark features are crucial elements for mapping and path planning. In the vSLAM literature, loop closure detection is a very important process for enhancing the robustness of the vSLAM algorithms. The most widely used algorithms for loop closure detection use a single descriptor. However, the performance of the single descriptors appears to worsen as the map keeps growing. One possible solution to this problem is to use multiple descriptors and combine them as in Naive and linear combinations. These approaches, however, have weaknesses in recognizing the correct locations due to overfitting and highbias, which hinder the generalization performance. This paper proposes the usage of ensemble learning to combine the predictions of multiple Bayesian filter models which make more accurate prediction than individual models. The proposed approach is validated on three public datasets; namely, Lip6 Indoor, Lip6 Outdoor and City Centre. The results show that the proposed ensemble algorithm significantly outperforms the single approaches with a recall of 80%, 97% and 87%, with 100% precision on the three datasets, and outperforms the Naive approach and the existing loop closure detection algorithms.

KW - Appearance-based localization

KW - Ensemble learning

KW - Loop closure detection

UR - http://www.scopus.com/inward/record.url?scp=85044176631&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85044176631&partnerID=8YFLogxK

M3 - Article

AN - SCOPUS:85044176631

VL - 25

SP - 235

EP - 242

JO - Pertanika Journal of Science and Technology

JF - Pertanika Journal of Science and Technology

SN - 0128-7680

IS - S6

ER -