RoadAnomalyDetectionThroughDeepLearningApproaches资源-CSDN文库

深度学习

需积分: 5 143 浏览量 2023-12-15 20:59:09 上传评论收藏 2.2MB PDF 举报

资源推荐

资源详情

资源评论

SPECIAL SECTION ON BIG DATA TECHNOLOGY AND

APPLICATIONS IN INTELLIGENT TRANSPORTATION

Received April 12, 2020, accepted June 17, 2020, date of publication June 24, 2020, date of current version July 6, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3004590

Road Anomaly Detection Through

Deep Learning Approaches

DAWEI LUO

1,2

, JIANBO LU

, AND GANG GUO

Department of Automotive Engineering, Chongqing University, Chongqing 400044, China

Research and Advanced Engineering, Ford Motor Company, Dearborn, MI 48124, USA

Corresponding author: Gang Guo (guogang@cqu.edu.cn)

This work was supported in part by the Chongqing Science and Technology Committee under Grant cstc2018jszx-cyzdX0074, and in part

by Ford Motor Company.

ABSTRACT This paper addresses road anomaly detection by formulating it as a classiﬁcation problem

and applying deep learning approaches to solve it. Besides conventional road anomalies, additional ones

are introduced from the perspective of a vehicle. In order to facilitate the learning process, the paper pays a

close attention to pattern representation, and proposes three sets of numeric features for representing road

conditions. Also, three deep learning approaches, i.e. Deep Feedforward Network (DFN), Convolutional

Neural Network (CNN), and Recurrent Neural Network (RNN), are considered to tackle the classiﬁcation

problem. The detectors, with respect to the three deep learning approaches, are trained and evaluated through

data collected from a test vehicle driven on various road anomaly conditions. The comparison study on the

detection performances is conducted by setting key hyper-parameters to certain sets of ﬁxed values. Also,

the comparison study on performances of each detector with respect to different pattern representations is

conducted. The results have shown the effectiveness of the proposed approaches and the efﬁciency of the

proposed feature representations in road anomaly detection.

INDEX TERMS Convolutional neural network, deep feedforward network, deep learning, pattern represen-

tation, recurrent neural network, road anomaly detection.

I. INTRODUCTION

Anomaly is something that deviates from what is standard or

the normal. Anomaly detection (AD) in real-world applica-

tions aims at determining if there are instances that are dra-

matically dissimilar to all the other instances [1]. A generic

form is the so-called outlier detection which refers to ﬁnding

patterns in data that do not conform to the expected normal

behavior [2]. A more concrete concept on an outlier is given

by [3], which is an observation that deviates so signiﬁcantly

from the other observations as to arouse suspicion that it was

generated by a different mechanism.

A rich body of literature [2]–[6] reviewed the approaches

for solving AD problems. As the advancement of deep

learning [7], generic methods and algorithms [1], [8]

have emerged by incorporating traditional AD approaches

with neural networks. This marriage has prompted various

novel AD methods that have been successfully applied to

The associate editor coordinating the review of this manuscript and

approving it for publication was Sabah Mohammed .

obstacles and anomalies detection in different application

domains [9]–[18].

In this paper, AD for road condition is of interest and

its main application is for intelligent transportation sys-

tems (ITS) and smart mobility. By using in-vehicle sensors

and their measurements, detecting anomalous road conditions

such as rough road, potholes, and speed bumps is pivotal

in ITS since those anomalies can cause ride discomfort and

potential damage to the vehicle’s chassis systems such as

wheel and tire assemblies, steering system, and suspension

systems. Furthermore, for an autonomous vehicle (AV) where

there are no human drivers to remember and monitor those

road conditions, such a lack of self-awareness could be a

limiting factor for its healthy operation. Hence, it might be

desirable to monitor the road anomalies encountered by the

vehicle itself. This information can be used as one of the

factors to schedule on-demand maintenance so as to pre-

vent unnecessary vehicle breakdown. This motivates us to

look into using vehicle as a ‘‘sensor’’ to detect the road

anomalies.

117390

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/

VOLUME 8, 2020

D. Luo et al.: Road AD Through Deep Learning Approaches

For road anomaly detection, unlike most of the research

works which focus on potholes and bumps only, we consider

a broader scope of anomalies. On the other hand, instead of

using range sensors or image sensors, raw data are acquired

by leveraging in-vehicle sensors. Those raw data are the

measurements of vehicle responses. Thus, road condition will

be inferred indirectly. Nowadays, lots of works focus on using

smartphone sensors, however, signatures of road anomalies

may be ﬁltered out due to the fact that the smartphones

are not rigidly ﬁxed on the vehicle. Moreover, as majority

of literatures rely on vertical signals, some complex types

of road anomalies can not be inferred except potholes or

bumps. For those cases, other vehicle response signals can

be used, together with inertial signals, to form multivariate

time series data. This will result in richer and more robust

feature representation, such that the modeling capabilities

of deep learning models can be fully explored. We tackle

road anomaly detection by formulating it as a classiﬁcation

problem and apply deep learning approaches by leveraging

multivariate time series data. The main contributions of this

paper are:

• The eight types of road (pavement) anomalies are spec-

iﬁed from the perspective of a vehicle.

• The three sets of numerical features for representing

road conditions are proposed. The comparison study on

the detection performances of a deep learning model

with respect to (w.r.t.) the three sets of features is con-

ducted.

• Road anomaly detection is formulated as a classiﬁca-

tion problem. A Deep Feedforward Network (DFN) is

applied to solve the classiﬁcation problem. In addi-

tion, a Convolutional Neural Network (CNN), and a

Recurrent Neural Network (RNN) are also applied.

Furthermore, comparison study on the detection per-

formance of the three approaches is conducted by

setting key hyper-parameters to certain pre-deﬁned

values.

The rest of this paper is organized as follows: section II

reviews the literatures that focus on road anomaly detec-

tion; section III describes the problem formulation of road

anomaly detection; section IV includes the data acquisi-

tion, pattern representation for road conditions, and dataset

construction; section V models road anomaly and presents

the architectures of the three deep learning approaches;

section VI shows the training results and compares the per-

formances of the three deep learning approaches; section VII

concludes the paper.

II. RELATED WORK

Road anomaly detection can be tackled by deﬁning and

modeling the reference (normal behavior) through physical

model based approaches [19], [20] and then ﬁnding the sim-

ilarity metric between the normal and the anomaly behav-

ior [21]. When the dynamical process of road anomaly is

characterized as low level of nonlinearity, free from time

delays, and less noisy in its observations, the physical model

based approaches are very efﬁcient. If the process, however,

is highly nonlinear or noisy, data driven approach might be

more efﬁcient and effective.

Road anomaly detectors based on support vector

machine (SVM) are studied [22]–[28], almost all of which

use smartphone sensors for data collection except [22].

Robust features of road anomaly are extracted by applying

signal processing techniques, such as wavelet decomposition,

resampling, and thresholding, to acceleration signal or other

types of inertial signals of vehicle responses. Besides SVM,

deep learning approaches also investigated in [29], [30].

Reference [29] applies a conventional DFN to learn a detector

by feeding with inertial signals acquired by both in-vehicle

sensors and smartphone sensors. Reference [30] applies a

unique deep learning algorithm which combines wavelets,

neural network, and Hilbert transform, to extract robust

features of an acceleration signal only. Moreover, [31] uses

unsupervised learning technique (clustering), to detect road

anomalies by analyzing the changes of driving behavior. For

most signal based road anomaly detection approaches, lots of

works rely on smartphone sensors. Thus, [32], [33] paid close

attention to implement a user-friendly framework for road

anomaly detection. Reference [32] focuses on developing

robust sensor measurements as well as sharing informa-

tion among users. Reference [33] proposes a smartphone

probe car (SPC) system to objectively assess and monitor

road conditions, which endows the system with awareness

of mounted posture of smartphones. It also performs lots

of signal processing in device level as well as in system

level.

For all the above research works, vehicle response data

acquired by either smartphone sensors or in-vehicle sen-

sors. Instead of using vehicle response signals, researchers

also take advantage of image data to perform road anomaly

detection directly [34]–[38]. Conventional image processing

techniques are applied extensively in those works. In addition,

machine learning approaches, such as SVM [36], special

designed CNN [37], and conventional CNN [38] are also

investigated and integrated with traditional image process

methods.

Although image based approaches can detect road anoma-

lies in advance, they may fail due to potential issues with

lighting and weather conditions. On the other hand, the smart-

phone sensors can pick up information which are not nec-

essarily related to vehicle responses and not all the vehicle

responses of the eight types of road anomalies deﬁned in this

paper can be inferred from the smartphone sensors. Those

motivate us to use in-vehicle sensors for data acquisition and

propose three sets of numerical features for pattern repre-

sentation of normal and anomalous road conditions. As few

research works apply CNN along with time series data in road

anomaly detection domain, we attempt to learn a CNN detec-

tor and compare its performance with DFN detector. Also,

there are few approaches based on RNN, we develop a RNN

detector which is capable of learning the complex dynamics

VOLUME 8, 2020 117391

D. Luo et al.: Road AD Through Deep Learning Approaches

FIGURE 1. Example plot of vehicle responses of a pothole test. The

patterns inside the ovals corresponds to the pothole negotiation.

of vehicle responses when encountering road anomalies, and

differentiate them from the normal road conditions. The per-

formance of the learned RNN detector is compared with the

other two detectors we investigated.

III. PROBLEM FORMULATION

Vehicle dynamic responses such as roll and yaw rate, acceler-

ations, and wheel speeds, can be sensed by existing in-vehicle

sensors. When it encounters anomalous road conditions such

as potholes and bumps, special patterns will show up in the

responses, which are captured by intervals of time sequences

through the measured signals. For example, Fig. 1 shows

time sequences of the measured signals and certain patterns

showing inside the ovals corresponds to a road anomaly.

This change of pattern from data point of view prompts us

using the responses to infer if a road anomaly has been

negotiated, so as to declare that the road conditions as an

anomaly. Instead of using physical model based methods,

deep learning approaches are proposed. More speciﬁcally,

the anomaly detection is formulated as a binary classiﬁcation

problem, which differentiates the normal road condition from

an anomalous road condition of any of the eight types in

set RA:

RA = {

pothole,

bump,

gravel,

cobblestone,

broken concrete,

curb impact,

the road condition that causes high wheel impact,

the road condition that causes severe vehicle body twist

All types of road anomalies in RA are used to describe the

status of anomalous road pavement. It is rarely to see road

anomalies, such as curb impact, vehicle body twist, and vehi-

cle wheel impact, showing up in literatures. Those three types

are introduced herein as we categorize road anomalies from

a vehicle’s perspective, which are unsatisfactory behaviors to

a vehicle.

Let X

= {x

(k)

∈ R

: k ∈ {1, 2, . . . , N }} be a

multivariate time series which is a sequence with length N .

It includes the collection of the raw samples corresponding

to the natural samples obtained through real-time sampling of

the selected m vehicle response signals, which are assumed to

be sampled at the same frequency. Obviously, the raw sample

(k)

is a m-tuple, and sampled at time instant k, which is also

regarded as a column vector in this paper:

(k)

= [x

(k)

, x

(k)

, . . . , x

(k)

]

, k ∈ {1, . . . , N } (1)

We deﬁne the ground truth label of x

(k)

= L

(k)

] (2)

where the labeling function L

will be addressed in

Section IV. Hence, the raw label set corresponds to X

follows:

= {y

(k)

: k ∈ {1, 2, . . . , N }} (3)

For time series data, sliding window techniques are usu-

ally used to divide the time-series into discrete segments in

order to reveal the underlying properties of its source. Hence,

a sliding window with ﬁxed length, TR, is applied to X

. Let

= {x

(i)

∈ R

m×TR

: i ∈ {0, 1, . . . , (N

− 1)}} be the set

resulting from applying the sliding window, whose elements

are the segments of X

. In this paper, we denote segment x

(i)

as the raw training sample. Thus, we refer TR as the training-

sample-to-raw-sample ratio since one (raw) training sample

for the learning models originates from TR consecutive raw

samples:

(i)

= [x

(i∗(TR−N

ors

)+1)

, x

(i∗(TR−N

ors

)+2)

. . . , x

(i∗(TR−N

ors

)+TR)

] (4)

where TR ≤ N , and N

ors

is the number of overlapped

raw samples between x

(i)

and x

(i+1)

. Therefore, the total

117392 VOLUME 8, 2020

剩余14页未读，继续阅读

评论收藏

内容反馈

就是菜啊！

粉丝: 54
资源: 2

Road Anomaly Detection Through Deep Learning Approaches

Deep Learning for Anomaly Detection A Review.pdf

Network Anomaly Detection-Machine Learning perspective

2.Deep Learning for Anomaly Detection A Review 论文分享（中）.pdf

Anomaly Detection_A Survey

A Discriminative Metric Learning Based Anomaly Detection Method

DEEP LEARNING FOR ANOMALY DETECTION A SURVEY中文翻译.zip

Network Anomaly Detection: A Machine Learning Perspective

Anomaly Detection as a Service: Challenges, Advances, and Opportunities

Unsupervised Anomaly Detection with GAN

Deep Learning: Practical Neural Networks with Java 完整高清英文azw3版

A Self-Adaptive Deep Learning-Based System for Anomaly Detection

Combining Filtering and Statistical Methods for Anomaly Detection

DEEP SEMI-SUPERVISED ANOMALY DETECTION

Online Learning and Sequential Anomaly Detection in Trajectories

Anomaly Detection in IP Networks

Practical Machine Learning_A New Look at Anomaly Detection-O'Reilly(2014).pdf

Forward Neural Network for Time Series Anomaly Detection

相关实用应用程序（Windows可用）

免费可用的ChatGPT网页版.zip

ChatGPT使用总结：150个ChatGPT提示词模板（完整版）

chromedriver-win64.zip

全国计算机二级WPSoffice精选350道选择题题库（含答案）.pdf

李飞飞自传 我看见的世界 The World I see

哈尔滨工业大学-ChatGPT调研报告-2023.3.6-94页.pdf

农村公交与异构无人机协同配送优化

4个亲测好用的ChatGPT4渠道

学术海报模板+论文科研+研究生

北森能力测评题库.zip

最新资源

李飞飞自传我看见的世界 The World I see