Uncertainty-Aware Lidar Place Recognition in Novel Environments

2025-04-15 0 0 1.76MB 8 页 10玖币

侵权投诉

Keita Mason1,2, Joshua Knights1,2, Milad Ramezani1, Peyman Moghadam1,2, Dimity Miller2

Abstract— State-of-the-art lidar place recognition models

exhibit unreliable performance when tested on environments

different from their training dataset, which limits their use in

complex and evolving environments. To address this issue, we

investigate the task of uncertainty-aware lidar place recognition,

where each predicted place must have an associated uncertainty

that can be used to identify and reject incorrect predictions.

We introduce a novel evaluation protocol and present the

ﬁrst comprehensive benchmark for this task, testing across

ﬁve uncertainty estimation techniques and three large-scale

datasets. Our results show that an Ensembles approach is

the highest performing technique, consistently improving the

performance of lidar place recognition and uncertainty estima-

tion in novel environments, though it incurs a computational

cost. Code is publicly available at https://github.com/csiro-

robotics/Uncertainty-LPR.

I. INTRODUCTION

Localisation is a crucial capability of mobile robots –

with an understanding of its location in a map, a robot

can navigate to new locations, monitor an environment, and

collaborate with other entities. Lidar place recognition (LPR)

algorithms use point clouds to enable robot localisation – a

robot can compare a recently captured point cloud with a

map of previous point clouds to identify its current loca-

tion, where the map can be generated on-the-ﬂy or ofﬂine.

These recognised locations, i.e., revisit areas, can be used as

loop closure constraints in a Simultaneous Localisation and

Mapping (SLAM) algorithm to mitigate drift, or provide re-

localisation in GPS-denied environments.

State-of-the-art approaches to LPR utilise deep neural

networks [1]–[6], and exhibit impressive localisation per-

formance when tested on environments included in the

training dataset [4]–[6]. However, when tested on a novel

environment,i.e., an environment not in the training dataset,

their performance drops substantially. As an example, a

MinkLoc3D model [4] can achieve 93% recall when trained

and tested on urban road scenes from the United Kingdom,

but this drops to 61% recall when tested on urban road scenes

from South Korea. This performance degradation highlights

a critical weakness of state-of-the-art LPR techniques (and

deep neural networks in general) – an inability to generalise

to conditions not represented in the training dataset.

Uncertainty-aware deep networks are an established ap-

proach for enabling reliable performance in novel condi-

tions [7]–[9]. Alongside each prediction, a network provides

1Authors are with the Robotics and Autonomous Systems, DATA61,

CSIRO, Brisbane, QLD 4069, Australia. 2Authors are with the School

of Electrical Engineering, Queensland University of Technology (QUT),

Brisbane, Australia.

Author contact emails: {joshua.knights, milad.ramezani,

peyman.moghadam}@data61.csiro.au, {kk.graves,

d24.miller}@qut.edu.au

Fig. 1: The bottom traversal shows the performance of MinkLoc3D

[4] when trained on Oxford road scenes, but tested on road scenes

from the Daejeon Convention Centre in South Korea. Tested on

this novel environment, MinkLoc3D predicts many false revisits. In

the top traversal, we show uncertainty-aware LPR with Ensembles,

where uncertainty is used to reject the 50% most uncertain queries.

an estimate of its uncertainty, where high uncertainty indi-

cates the network is more likely to make a mistake. While

uncertainty-aware deep networks have been explored in many

computer vision ﬁelds and robotics applications [9]–[18], no

existing research explores uncertainty estimation for LPR.

In this paper, we introduce uncertainty-aware lidar place

recognition and present a comprehensive benchmark of this

task. Our contributions are as follows:

1) We formalise the uncertainty-aware LPR task for the

prediction of lidar-based localisation failure (Sec. III).

2) Drawing inspiration from state-of-the-art techniques in

related ﬁelds, we implement one standard baseline and

four uncertainty-aware baselines for the LPR setting

(Sec. III-C).

3) We introduce an evaluation protocol that utilises three

large-scale datasets and a range of metrics to quantify

place recognition ability and uncertainty estimation for

LPR in novel environments (Sec. IV).

4) We analyse the performance of all baseline methods on

this evaluation protocol, exploring how different error

types inﬂuence performance, as well as computational

cost (Sec. V).

We hope that this work enables and stimulates further

research into this important area by providing an extensive

evaluation protocol and initial benchmark for comparison.

II. RELATED WORK

Before introducing uncertainty-aware LPR, we ﬁrst con-

textualise the work by reviewing the existing state-of-the-art

in LPR, and then discussing existing research into uncer-

tainty estimation in retrieval tasks.

arXiv:2210.01361v3 [cs.CV] 12 Jul 2023

A. Lidar Place Recognition

LPR utilising 3D point clouds has been signiﬁcantly ex-

plored in the last few years. LPR approaches identify similar

places (revisited areas) by encoding high-dimensional point

clouds into discriminative embeddings (often referred to as

descriptors). Handcrafted LPR methods [19]–[23] generate

local descriptors by segmenting point clouds into patches, or

global descriptors that show the relationship between all the

points in a point cloud.

Recent state-of-the-art LPR approaches have been dom-

inated by deep learning-based architectures due to their

impressive performance [2]–[6], [24]–[28]. These approaches

typically utilise a backbone architecture to extract local

features from the point cloud, which are then aggregated into

a global descriptor. The speciﬁc design of these components

varies signiﬁcantly between different works; PointNet [24],

graph neural networks [3], transformers [2], [6], [25], and

sparse-voxel convolutional networks [4]–[6], [26] have all

been proposed as local feature extractors, and aggregation

methods include NetVLAD [29], Generalised Mean Pooling

(GeM) [30] and second-order pooling [5], [31].

B. Uncertainty Estimation for Retrieval Tasks

Though there are a number of works exploring uncertainty

estimation in lidar object detection [15], [17], [18], [32],

[33] and point cloud segmentation [16], no existing works

explore uncertainty estimation for LPR. While recent work

by Knights et al. [34] shares a similar motivation to our work

– reliable performance in novel environments – they explore

incremental learning and speciﬁcally the issue of catastrophic

forgetting.

Image retrieval is a ﬁeld of computer vision that shares

a similar problem setup to LPR (though notably operat-

ing on images rather than point clouds). When estimat-

ing uncertainty for image retrieval, recent works learn an

uncertainty estimate by adding additional heads to their

network architecture [13], [14], [35]. Shi et al. [13] examine

uncertainty-aware facial recognition, where face embeddings

are modelled as Gaussian distributions by learning both a

mean vector and variance vector.

Warburg et al. [35] follow a similar approach, introducing

a ‘Bayesian Triplet Loss’ to extend training to also include

negative probabilistic embeddings. Most recently, STUN [14]

was proposed for uncertainty-aware visual place recognition.

STUN presents a student-teacher paradigm to learn a mean

vector and variance vector, using the average variance to

represent uncertainty [14]. Given the high performance of

these approaches in the related image retrieval task, we adapt

several of these methods to the LPR setting to serve as

baselines for our benchmark.

III. METHODOLOGY

We ﬁrst deﬁne the LPR task, and then formalise

uncertainty-aware LPR. Following this, we introduce the

baseline methods used for our benchmark.

A. Lidar Place Recognition

During LPR evaluation, a database contains point clouds

with attached location information. This database can be a

previously curated map or can be collected online as an

agent explores an environment. Given a query, i.e., a new

point cloud from an unknown location, an LPR model must

localise the query by ﬁnding the matching point cloud in

the database. If the predicted database location is within a

minimum global distance to the true query location, the pre-

diction is considered correctly recalled. In this conﬁguration,

LPR performance is evaluated from the average recall of all

tested queries [1], [4], [6].

We are motivated by the observation that perfect recall

in an LPR setting does not currently exist, and may not be

attainable in some applications – when operating in dynamic

and evolving environments, or dealing with sensor noise,

the potential for error always exists. In this case, we argue

that LPR models should additionally be able to estimate

uncertainty in their predictions, i.e.,know when they don’t

know. We formalise this below as uncertainty-aware LPR.

B. Uncertainty-aware Lidar Place Recognition

In uncertainty-aware LPR, each predicted match between a

query and database entry should be accompanied by a scalar

uncertainty estimate U. This uncertainty represents the lack

of conﬁdence in a predicted location.

Following the existing LPR setup, the primary goal in

uncertainty-aware LPR is to maximise correct localisations

(i.e., recall). Uncertainty-aware LPR extends on this by addi-

tionally requiring models to identify incorrect predictions by

associating high uncertainty. We formulate this as a binary

classiﬁcation problem, where Uis compared to a decision

threshold λto classify whether an LPR prediction is correct

or incorrect:

Fλ(U) = Correct, U ≤λ

Incorrect, U > λ . (1)

Incorrect predictions can arise for two reasons: (1) the

query is from a location that is not present in the database,

or (2) the query is from a location in the database, but the

LPR model selects the incorrect database match. We refer to

these two error types as ‘no match error’ and ‘incorrect match

error’ respectively, and analyse them in detail in Sec. V-B.

C. Baseline Approaches

To benchmark uncertainty-aware LPR, we adapt a number

of uncertainty estimation techniques existing in related ﬁelds

to the LPR setting.

Standard LPR Network: As explored in Sec. II-A, state-

of-the-art LPR techniques utilise a deep neural network to

reduce a point cloud to a descriptor d. Given a database

of Nprevious point clouds and locations, a standard LPR

network converts this to a database Dof N L-dimensional

descriptors, D={di∈RL}N

i=1. During evaluation, a query

point cloud Pq∈RM×3, with Mpoints, is reduced to a

query descriptor dq∈RL. This query descriptor is compared

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Uncertainty-AwareLidarPlaceRecognitioninNovelEnvironmentsKeitaMason1,2,JoshuaKnights1,2,MiladRamezani1,PeymanMoghadam1,2,DimityMiller2Abstract—State-of-the-artlidarplacerecognitionmodelsexhibitunreliableperformancewhentestedonenvironmentsdifferentfromtheirtrainingdataset,whichlimitstheiruseincomplex...

展开>> 收起<<

Uncertainty-Aware Lidar Place Recognition in Novel Environments.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Uncertainty-Aware Lidar Place Recognition in Novel Environments

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: