ConvSequential-SLAM: A Sequence-Based, Training-Less Visual Place Recognition Technique for Changing Environments

Mihnea-Alexandru Tomia,Klaus D Mcdonald-Maier,Shoaib Ehsan,Mubariz Zaffar,Michael J Milford

doi:10.1109/access.2021.3107778

Mihnea-Alexandru Tomia, Klaus D Mcdonald-Maier + Show 3 more

Open Access

https://doi.org/10.1109/access.2021.3107778

Copy DOI

Abstract

Visual Place Recognition (VPR) is the ability to correctly recall a previously visited place under changing viewpoints and appearances. A large number of handcrafted and deep-learning-based VPR techniques exist, where the former suffer from appearance changes and the latter have significant computational needs. In this paper, we present a new handcrafted VPR technique, namely ConvSequential-SLAM, that achieves state-of-the-art place matching performance under challenging conditions. We utilise sequential information and block-normalisation to handle appearance changes, while using regional-convolutional matching to achieve viewpoint-invariance. We analyse content-overlap in-between query frames to find a minimum sequence length, while also re-using the image entropy information for environment-based sequence length tuning. State-of-the-art performance is reported in contrast to 9 contemporary VPR techniques on 4 public datasets. Qualitative insights and an ablation study on sequence length are also provided.

Highlights

Visual Place Recognition (VPR) is the ability of a robot to correctly identify a previously visited place using visual information
This approach can only be used by this particular algorithm presented in the paper, and cannot be applied to all possible algorithms; 4) Building upon the sequence length generated by analysing consecutive query images, we use the entropy computation for salient region extraction to formulate an optimal dynamic sequence length, instead of a constant sequence length, as used in sequencebased VPR techniques
We discuss results from a place matching encoding time for dynamic k will vary between the lowest performance point, in terms of accuracy, AUC-PR and PCU. value

Summary

INTRODUCTION

VPR is the ability of a robot to correctly identify a previously visited place using visual information. We propose a novel sequence-based and training-free VPR technique, namely ConvSequentialSLAM, that is successfully able to perform Visual Place Recognition (VPR) under changing viewpoint and appearance conditions. We developed an analysis based on information-gain from consecutive query images to determine the minimum sequence length needed This approach can only be used by this particular algorithm presented in the paper, and cannot be applied to all possible algorithms; 4) Building upon the sequence length generated by analysing consecutive query images, we use the entropy computation for salient region extraction to formulate an optimal dynamic sequence length, instead of a constant sequence length, as used in sequencebased VPR techniques. When the Information Gain module provides its best sequence length (e.g. Inf o Gain < IT ), we proceed to calculate the sequential entropy (see subsection III-F) for that sequence of query images and determine whether this has the optimal length

ENTROPY MAP AND ROI EXTRACTION

RESULTS

ACCURACY present the PCU value of our technique against other VPR

CONCLUSIONS AND FUTURE DIRECTIONS