Vehicle Localization using 76GHz Omnidirectional Millimeter-Wave Radar for Winter Automated Driving.pdf

海尔

177

7页

0次

2021-02-23

50墨值下载

Vehicle Localization using 76GHz Omnidirectional Millimeter-Wave

Radar for Winter Automated Driving*

Keisuke Yoneda

, Naoya Hashimoto

, Ryo Yanase

, Mohammad Aldibaja

and Naoki Suganuma

Abstract— This paper presents the 76GHz MWR (Millimeter-

Wave Radar)-based self-localization method for automated

driving during snowfall. Previously, there were many LIDAR

(Light Detection and Ranging)-based localization techniques

proposed for high measurement accuracy and robustness to

changes of day and night. However, they did not provide effec-

tive approaches for snow conditions because of sensing noise

(i.e., environmental resistance) created by snowfall. Therefore,

this paper developed a MWR-based map generation and a

real-time localization method by modeling the uncertainties of

error propagation. Quantitative evaluations are performed on

driving data with and without snow conditions, using a LIDAR-

based method as the baseline. Experimental results show that a

lateral root mean square error of about 0.25m can be obtained,

regardless of the presence or absence of snowfall. Thus, it can

be investigated that a potential performance of radar-based

localization.

I. INTRODUCTION

Automated vehicle technology has reached the era of

comprehensive testing and practical applications. Many auto-

motive companies and research organizations have conducted

driving experiments on public roads [1], [2]. Such vehicles

are equipped with various sensors including LIDAR (Light

Detection and Ranging), MWR (Millimeter-Wave Radar), the

camera and GNSS/IMU (Global Navigation Satellite System

/Inertial Measurement Unit) system to percept surroundings

and allow autonomous behavior. These sensor data are used

to attain the following objectives.

• Environmental perception: detect static/dynamic object

and dynamic road object (e.g., trafﬁc signal status).

• Self-localization: estimate own position on a precise

digital map.

• Motion planning: generate safety trajectory, considering

trafﬁc rules.

• Motion control: determine adequate control signals for

steering, acceleration, and braking.

In order to achieve safe driving on a public road, a

common approach is implementing robust perception and

decision-making systems using precise digital maps. For

example, by referring to the map relating to trafﬁc light

position, it is possible to accurately and quickly recognize

the state of the trafﬁc light [3], [4]. To smoothly control the

vehicle by referring to the map according to the estimated

position, decimeter-level positional accuracy is required. The

*This work was supported by Kanazawa University

K. Yoneda, N. Hashimoto, R. Yanase, M. Aldibja and N. Suganuma are

with Kanazawa University, Kakuma-cho, Kanazawa, Ishikawa, 920-1192,

Japan.

k.yoneda@staff.kanazawa-u.ac.jp

general strategy of self-localization is map-matching using a

precisely predeﬁned map.

LIDAR-based methods have been proposed because of

their high measurement accuracy and robustness to changes

in day and night. In [5], a method was proposed in which

a map was created by mapping the infrared reﬂectivity of

the road surface. The vehicle position was then estimated

as a probability distribution by applying template-matching

between the map and the real-time LIDAR point cloud. This

approach was implemented on road-paint information, and

reported positional accuracy at the decimeter-level. However

the problem with the LIDAR-based method is the bad

weather, such as rain and snowfall. For example, during

snowfall, landmarks cannot be observed because the road

surface is occluded. In order to resolve such difﬁculties,

an algorithm for reconstructing observation information of

LIDAR was proposed for the situation in which the road

surface is partially occluded [6]. However, it can not be

applied to situations where the road surface is completely

occluded. Methods have also been proposed that exploit the

features of the surrounding buildings by utilizing the 3-D

point cloud map [7], [8]. However, because the shape of

the roadside changes during snowfall, it is not an effective

approach for snow conditions. Therefore, MWR can be used

as a sensor to robustly observe surrounding objects during

snowfall.

MWR is excellent for penetrating environmental resis-

tance. However, it has the disadvantage of sparse observation

information, owing to the low resolution of the angular

direction compared to LIDAR. In previous studies, the self-

localization methods were proposed to utilize MWR [9],

[10]. In [9], a map generation method was proposed that

clustered radar observations. Vehicle position was estimated

using a particle ﬁlter. However, sufﬁcient positioning accu-

racy was not obtained for self-driving because the obtained

accuracy was the meter-level. In [10], a method using MWR

and images from the around view monitor (AVM) was

proposed. Road-level localization was implemented using

MWR and then the accurate vehicle position is estimated

using the AVM image by matching lane boundaries as a lane-

level localization. Even with this method, it was assumed

that sufﬁcient precision could not be obtained with only

MWR. Furthermore, the lane-level localization could not

be utilized when lane boundaries were occluded by snow.

Therefore, this study proposes a self-localization method

using 76GHz MWR observation, measuring the performance

in snow conditions and comparing with the LIDAR-based

approach. The evaluation is performed using a high-precision

2018 IEEE Intelligent Vehicles Symposium (IV)

Changshu, Suzhou, China, June 26-30, 2018

IMU

Rotary Encoder Millimeter-wave Radar

3-D LIDAR

Frontal Camera

GNSS Antenna

(a) Experimental Vehicle

(b) Omni-directional Millimeter-wave Radar

Fig. 1. Experimental Vehicle

RTK-GNSS (Real Time Kinematic GNSS) with a post-

processing, providing centimeter positional accuracy. The

key contributions of this paper are as follows.

• Development of both map generation and localization

by modeling uncertainties of MWR, based on error

propagation.

• Quantitative evaluations for accuracy of self-localization

against snowfall, comparing with LIDAR-based method

as baseline.

The rest of this paper is composed as follows. Section II

introduces the experimental vehicle. Section III explains the

2-D map generation using MWR. The proposed localization

method is detailed in Section IV. Section V describes numeri-

cal experiments for urban driving data and evaluation results.

Finally, Section VI concludes with the obtained results and

offers directions for future work.

II. EXPERIMENTAL VEHICLE

Figure 1(a) illustrates our experimental vehicle. The vehi-

cle is equipped with an Applanix POS/LV220 coupled GNSS

and IMU system. It provides position (i.e., latitude, longitude

and altitude) and orientation (pitch, yaw, roll) at 100Hz.

A 3-D LIDAR Velodyne HDL-64E S2 with 64 separate

beams is mounted on a vehicle to take measurements of the

environment. It measures 3-D omnidirectional distance under

a frequency of about 10Hz. Nine MWRs are installed inside

the front and rear bumpers to recognize distant objects, as

shown in Fig. 1(b). It measures distance, angle and relative

velocity for objects at 20Hz. The horizontal ﬁeld-of-view is

40deg. The number of observations for each MWR is 40.

Map Generation

2D Radar Image Map

RTK-GNSS Radar GNSS/IMU

Moving Object

Tracking

Static Object

Extraction

Mapping &

Updating Radar Map

Realtime Self-Localization

Generating

Obs. Image

Template

Matching

Probability

Updating

Output

Estimated Pose

Dead-

Reckoning

Offset

Updating

Fig. 2. Proposed Localization Method

III. MAP GENERATION

A. Method

A reference map is generated by mapping MWR ob-

servations using the RTK-GNSS with post-processing. The

measuring angle sensing accuracy of 76GHz MWR is gen-

erally not precise compared to LIDAR and 79GHz MWR.

Therefore, it is necessary to consider measurement accuracy

during the mapping procedure. According to the left-side of

Fig. 2, the reference map is generated using the following

processes.

1) Object tracking: estimate static/dynamic bbjects.

2) Static object extraction: remove dynamic objects.

3) Mapping: update a probability of the static objects in

each map pixel.

B. Object Tracking

Object tracking estimates static/dynamic objects us-

ing MWR observations. The interactive multiple model

(IMM) [11] is adopted to integrate multiple types of motions

to estimate position, velocity, and acceleration for the object.

Constant acceleration, constant velocity, and stop models

are deﬁned as motion models. The state variable, x

mwr

deﬁned in Eq. (1).

mwr

= [p

, p

, v

, a

]

(1)

mwr

= [w

˙p

, w

˙p

, w

˙a

, w

˙a

]

(2)

where, p

and p

are object position in global coordinates

of the Universal Transverse Mercator (UTM) [12]. v

, a

, and a

are velocity and acceleration for the x and y

coordinates, respectively. w

mwr

is a process noise vector for

corresponding variables. For example, a state equation of the

independent constant acceleration model can be formulated

972

Vehicle Localization using 76GHz Omnidirectional Millimeter-Wave

Radar for Winter Automated Driving*

Keisuke Yoneda

, Naoya Hashimoto

, Ryo Yanase

, Mohammad Aldibaja

and Naoki Suganuma

Abstract— This paper presents the 76GHz MWR (Millimeter-

Wave Radar)-based self-localization method for automated

driving during snowfall. Previously, there were many LIDAR

(Light Detection and Ranging)-based localization techniques

proposed for high measurement accuracy and robustness to

changes of day and night. However, they did not provide effec-

tive approaches for snow conditions because of sensing noise

(i.e., environmental resistance) created by snowfall. Therefore,

this paper developed a MWR-based map generation and a

real-time localization method by modeling the uncertainties of

error propagation. Quantitative evaluations are performed on

driving data with and without snow conditions, using a LIDAR-

based method as the baseline. Experimental results show that a

lateral root mean square error of about 0.25m can be obtained,

regardless of the presence or absence of snowfall. Thus, it can

be investigated that a potential performance of radar-based

localization.

I. INTRODUCTION

Automated vehicle technology has reached the era of

comprehensive testing and practical applications. Many auto-

motive companies and research organizations have conducted

driving experiments on public roads [1], [2]. Such vehicles

are equipped with various sensors including LIDAR (Light

Detection and Ranging), MWR (Millimeter-Wave Radar), the

camera and GNSS/IMU (Global Navigation Satellite System

/Inertial Measurement Unit) system to percept surroundings

and allow autonomous behavior. These sensor data are used

to attain the following objectives.

• Environmental perception: detect static/dynamic object

and dynamic road object (e.g., trafﬁc signal status).

• Self-localization: estimate own position on a precise

digital map.

• Motion planning: generate safety trajectory, considering

trafﬁc rules.

• Motion control: determine adequate control signals for

steering, acceleration, and braking.

In order to achieve safe driving on a public road, a

common approach is implementing robust perception and

decision-making systems using precise digital maps. For

example, by referring to the map relating to trafﬁc light

position, it is possible to accurately and quickly recognize

the state of the trafﬁc light [3], [4]. To smoothly control the

vehicle by referring to the map according to the estimated

position, decimeter-level positional accuracy is required. The

*This work was supported by Kanazawa University

K. Yoneda, N. Hashimoto, R. Yanase, M. Aldibja and N. Suganuma are

with Kanazawa University, Kakuma-cho, Kanazawa, Ishikawa, 920-1192,

Japan.

k.yoneda@staff.kanazawa-u.ac.jp

general strategy of self-localization is map-matching using a

precisely predeﬁned map.

LIDAR-based methods have been proposed because of

their high measurement accuracy and robustness to changes

in day and night. In [5], a method was proposed in which

a map was created by mapping the infrared reﬂectivity of

the road surface. The vehicle position was then estimated

as a probability distribution by applying template-matching

between the map and the real-time LIDAR point cloud. This

approach was implemented on road-paint information, and

reported positional accuracy at the decimeter-level. However

the problem with the LIDAR-based method is the bad

weather, such as rain and snowfall. For example, during

snowfall, landmarks cannot be observed because the road

surface is occluded. In order to resolve such difﬁculties,

an algorithm for reconstructing observation information of

LIDAR was proposed for the situation in which the road

surface is partially occluded [6]. However, it can not be

applied to situations where the road surface is completely

occluded. Methods have also been proposed that exploit the

features of the surrounding buildings by utilizing the 3-D

point cloud map [7], [8]. However, because the shape of

the roadside changes during snowfall, it is not an effective

approach for snow conditions. Therefore, MWR can be used

as a sensor to robustly observe surrounding objects during

snowfall.

MWR is excellent for penetrating environmental resis-

tance. However, it has the disadvantage of sparse observation

information, owing to the low resolution of the angular

direction compared to LIDAR. In previous studies, the self-

localization methods were proposed to utilize MWR [9],

[10]. In [9], a map generation method was proposed that

clustered radar observations. Vehicle position was estimated

using a particle ﬁlter. However, sufﬁcient positioning accu-

racy was not obtained for self-driving because the obtained

accuracy was the meter-level. In [10], a method using MWR

and images from the around view monitor (AVM) was

proposed. Road-level localization was implemented using

MWR and then the accurate vehicle position is estimated

using the AVM image by matching lane boundaries as a lane-

level localization. Even with this method, it was assumed

that sufﬁcient precision could not be obtained with only

MWR. Furthermore, the lane-level localization could not

be utilized when lane boundaries were occluded by snow.

Therefore, this study proposes a self-localization method

using 76GHz MWR observation, measuring the performance

in snow conditions and comparing with the LIDAR-based

approach. The evaluation is performed using a high-precision

2018 IEEE Intelligent Vehicles Symposium (IV)

Changshu, Suzhou, China, June 26-30, 2018

IMU

Rotary Encoder Millimeter-wave Radar

3-D LIDAR

Frontal Camera

GNSS Antenna

(a) Experimental Vehicle

(b) Omni-directional Millimeter-wave Radar

Fig. 1. Experimental Vehicle

RTK-GNSS (Real Time Kinematic GNSS) with a post-

processing, providing centimeter positional accuracy. The

key contributions of this paper are as follows.

• Development of both map generation and localization

by modeling uncertainties of MWR, based on error

propagation.

• Quantitative evaluations for accuracy of self-localization

against snowfall, comparing with LIDAR-based method

as baseline.

The rest of this paper is composed as follows. Section II

introduces the experimental vehicle. Section III explains the

2-D map generation using MWR. The proposed localization

method is detailed in Section IV. Section V describes numeri-

cal experiments for urban driving data and evaluation results.

Finally, Section VI concludes with the obtained results and

offers directions for future work.

II. EXPERIMENTAL VEHICLE

Figure 1(a) illustrates our experimental vehicle. The vehi-

cle is equipped with an Applanix POS/LV220 coupled GNSS

and IMU system. It provides position (i.e., latitude, longitude

and altitude) and orientation (pitch, yaw, roll) at 100Hz.

A 3-D LIDAR Velodyne HDL-64E S2 with 64 separate

beams is mounted on a vehicle to take measurements of the

environment. It measures 3-D omnidirectional distance under

a frequency of about 10Hz. Nine MWRs are installed inside

the front and rear bumpers to recognize distant objects, as

shown in Fig. 1(b). It measures distance, angle and relative

velocity for objects at 20Hz. The horizontal ﬁeld-of-view is

40deg. The number of observations for each MWR is 40.

Map Generation

2D Radar Image Map

RTK-GNSS Radar GNSS/IMU

Moving Object

Tracking

Static Object

Extraction

Mapping &

Updating Radar Map

Realtime Self-Localization

Generating

Obs. Image

Template

Matching

Probability

Updating

Output

Estimated Pose

Dead-

Reckoning

Offset

Updating

Fig. 2. Proposed Localization Method

III. MAP GENERATION

A. Method

A reference map is generated by mapping MWR ob-

servations using the RTK-GNSS with post-processing. The

measuring angle sensing accuracy of 76GHz MWR is gen-

erally not precise compared to LIDAR and 79GHz MWR.

Therefore, it is necessary to consider measurement accuracy

during the mapping procedure. According to the left-side of

Fig. 2, the reference map is generated using the following

processes.

1) Object tracking: estimate static/dynamic bbjects.

2) Static object extraction: remove dynamic objects.

3) Mapping: update a probability of the static objects in

each map pixel.

B. Object Tracking

Object tracking estimates static/dynamic objects us-

ing MWR observations. The interactive multiple model

(IMM) [11] is adopted to integrate multiple types of motions

to estimate position, velocity, and acceleration for the object.

Constant acceleration, constant velocity, and stop models

are deﬁned as motion models. The state variable, x

mwr

deﬁned in Eq. (1).

mwr

= [p

, p

, v

, a

]

(1)

mwr

= [w

˙p

, w

˙p

, w

˙a

, w

˙a

]

(2)

where, p

and p

are object position in global coordinates

of the Universal Transverse Mercator (UTM) [12]. v

, a

, and a

are velocity and acceleration for the x and y

coordinates, respectively. w

mwr

is a process noise vector for

corresponding variables. For example, a state equation of the

independent constant acceleration model can be formulated

972

of 7

50墨值下载

database

关注

评论