3D Mapping of Indoor Parking Space Using Edge Consistency

Census Transform Stereo Odometry

Junesuk Lee and Soon-Yong Park

a

School of Electronic and Electrical Engineering, Kyungpook National University, Daegu, South Korea

Keywords: 3D Mapping, 3D Reconstruction, 3D Scanning, Visual Odometry, Parking Ramp, Parking Space.

Abstract: In this paper, we propose a real-time 3D mapping system for indoor parking ramps and spaces. Visual

odometry is calculated by applying the proposed Edge Consistency Census Transform (ECCT) stereo

matching method. ECCT works strongly in repeated patterns and reduces drift errors in the vertical direction

of the ground caused by Kanade-Lucas-Tomasi stereo matching of VINS-FUSION algorithm. We propose a

mobile mapping system that uses a stereo camera and 2D lidar for data set acquisition. The parking ramp and

spaces dataset are obtained using the mobile mapping system and are reconstructed using the proposed system.

The proposed system performs the 3D mapping of the parking ramp and spaces dataset that is obtained using

the mobile mapping system. We present the error of the normal vector with respect to the ground of the

parking space as a quantitative evaluation for performance comparison with the previous method. Also, we

present 3D mapping results as qualitative results.

1 INTRODUCTION

With the rapid development of technologies related to

autonomous driving, the 3D mapping technology of

real space is being actively researched. Self-driving

cars use the HD map built in advance to safely drive

in urban and highway environments, estimate the

location of the car, plan the driving route, and safely

drive to the destination (Kim et al., 2021; Ding et al.,

2021). To create an HD map, the vehicle is combined

with the camera, lidar, IMU, and GPS sensor to scan

the surrounding environment. Such a device is called

Mobile Mapping System (MMS) (Roh et al., 2016).

In addition, the HD map is provided information such

as traffic lights, traffic signs, and lanes that can affect

the localization and path planning of autonomous

vehicles (Elhousni et al., 2020). Considering the

complete driving of the autonomous vehicle, the

vehicle must be driven from the parked location to the

parking spaces of the destination. Autonomous valet

parking research (Qin et al., 2020; Chirca et al., 2015)

aims to park a vehicle at a certain spot in the parking

spaces. In addition to these studies, the problem of

entering the parking space inside the building from

the outdoor environment should be considered. In

most cases of underground or above-ground parking

a

https://orcid.org/ 0000-0001-5090-9667

included inside a building, it is necessary to pass

through the parking ramp. In general, it is difficult to

perform localization in the computer vision field

because the parking ramp section has few textures

and similar structures. Especially, in the case of a

parking ramp in a building, if the width is narrow and

the illumination is low, sit is difficult for autonomous

vehicles to enter. Therefore, as one method for safe

driving on the parking ramp, there is providing

information on the width, height, length, curvature,

and slope of the ramp section. In the International

Building Code (International Code Council, 2018),

(a) Parking ramp (b) Parking space

Figure 1: Experimental results of the 3D reconstruction for

the parking ramp and space by the proposed 3D mapping

system.

Lee, J. and Park, S.

3D Mapping of Indoor Parking Space Using Edge Consistency Census Transform Stereo Odometry.

DOI: 10.5220/0011789100003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages

1015-1020

ISBN: 978-989-758-634-7; ISSN: 2184-4321

Copyright

c

2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

1015

Figure 2: System overview of 3D reconstruction for the parking ramp and parking spaces.

the building law for parking ramp sections suitable

for automobile facilities have been established.

International building codes include width, height,

length, curvature, and slope for ramp sections of

buildings. To measure this information, 3D spatial

scanning must be accompanied.

In this paper, we propose a 3D mapping system

for parking ramps and parking spaces and a mobile

mapping device composed of a stereo camera and 2D

lidar. The 3D mapping system calculates visual

odometry using the proposed ECCT stereo-matching

method. The overall system structure of visual

odometry is based on VINS-FUSION (Qin et al.,

2019). In the parking spaces mapping process, VINS-

FUSION's Kanade-Lucas-Tomasi (KLT) stereo

matching method causes a large drift error in the

vertical direction of the ground (Bouguet, 1999). To

reduce this drift error, we estimate the visual

odometry using the proposed ECCT stereo matching.

ECCT stereo matching is robust to repeated patterns

and symmetrical patterns because it considers the

continuity and rotation of the edges for the matching

block. Next, the point data of the 2D lidar is projected

into the 3D space using the 3D coordinate

transformation matrix between the stereo camera and

the 2D Lidar sensor. Fig. 1 shows the 3D mapping

result generated by projecting 2D lidar data into 3D

space.

2 SYSTEM OVERVIEW

The proposed 3D mapping device configuration is

shown in Fig. 3. The device consists of a stereo

camera and 2D lidar, the stereo camera acquires

synchronized data through a hardware trigger, and the

2D lidar data acquire synchronized data through

software synchronization with the stereo camera. And

the system overview of the real-time 3D mapping is

shown in Fig. 2. This system takes stereo images as

inputs and estimates stereo-based visual odometry

(VO). The reference coordinate system of the

estimated visual odometry is the left camera

coordinate system. In the experiments, we found that

VINS-FUSION had a drift problem in the vertical

direction of the ground as shown in Fig. 4. VINS-

FUSION performs stereo matching based on the KLT

feature tracker. We improve the drift problem in the

vertical direction of the ground by applying the

proposed ECCT stereo matching method. The 2D

lidar data is projected onto the 3D space through

coordinate system transformation for the estimated

VO.

Proposed ing device 3D reconstruction of the parking space

Figure 3: Proposed 3D mapping device for the parking

ramp and parking space.

Figure 4: The drift problem with respect to the vertical

direction of the ground. The red point cloud is the result of

using KLT-based stereo matching. The green point cloud is

the result of using the proposed ECCT stereo-matching

method.

2D Lida

r

Camera

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

1016

3 STEREO BLOCK MATCHING

We apply the spare stereo matching method

considering real-time applications. A stereo matching

point is searched for the tracked feature points

between the previous frame and the current frame.

The KLT method is used to trace the feature points

between the previous frame and the current frame.

We propose ECCT stereo-matching for stereo block

matching. This method extends the traditional census

transform (Zabih and Woodfill, 1994). The proposed

method assumes non-rectified stereo images.

Rectified stereo images can be easily searched using

the epipolar line range search. However, a large part

of the image is removed in the process of creating a

rectified stereo image. Therefore, the field of view

advantage is lost when used in wide-angle images

(Lv, 2017). We use a wide-angle camera to take

advantage of the wide-angle FOV. We use a window

kernel of size n n for stereo matching, where n is

an odd value of three or greater. The left image is used

as a reference image and the search for stereo

matching points in the right image. The search range

searches 𝑤

ℎ

from the left image to the

reference pixel position of the right image. After

searching all pixels in the search range, the pixel point

with the lowest matching cost value is selected as the

stereo matching point. In this paper, n is three and

𝑤

ℎ

is 21 7.

3.1 Census Transform

Census Transform (CT) creates a binary pattern of 0

and 1 in a window block by comparison between the

central pixel and the surrounding pixels. Generates 1

if the center pixel is greater than the neighboring

pixel, and 0 if it is less than the neighboring pixel.

Then, the matching cost is calculated through the

Hamming distance (Liu and Na, 2022) of the

generated binary pattern. However, Census

Transform Cost (CTC) can produce the same cost

results despite different patterns within the search

area. This problem often occurs in the repeated

pattern (Liu and Collins, 2000) area. Dense stereo

matching compensates for this problem mostly in the

depth refinement stage (Lee et al., 2012). However,

our system must perform spare stereo matching while

satisfying real-time performance. Fig. 5 shows an

example of the case where the same CTC exists in a

size with a window size of 3. If multiple blocks with

the same matching cost are found within the search

area, matching blocks can be selected in two ways.

The first is to randomly select among the same

matching cost blocks. The second is to select a

matching cost block suitable for the constraint by

adding a constraint. We adopt the method of selecting

the final matching point by adding a constraint.

Figure 5: Example of the equal Census Transform cost for

different patterns.

3.2 Edge Consistency Census

Transform

We propose ECCT stereo-matching to improve CT

stereo-matching. We are motivated by fast features

(Rosten et al., 2008). The flowchart of the proposed

method is shown in Fig. 2. The consistency of the

edge area of the binary pattern generated by CT is

checked. The edge in the window means the yellow

cell in Fig. 6. The order of calculating Edge

Consistency Census Transform Cost (ECCTC) is as

follows: First, the maximum number of consecutive

zeros is searched from the binary bit calculated by CT

and defined as the Max count of edge consistency.

Assuming that the starting point and the ending point

are connected, search for the number of consecutive

zeros in clockwise order as shown in Fig. 6. Then, the

difference between the total number of edge cells and

the max count of edge consistency is calculated. The

total number of edge cells (𝑅

) is defined by (1).

Fig. 7 shows an example of calculating ECCTC. In

the 1st row of Fig. 7, The max count of edge

consistency in the binary bit is 5. And 𝑅

is 8,

ECCTC is 8-5=3. Similarly, in the second row of Fig.

7, the max count of edge consistency of binary bits is

2 and ECCTC is 8-2=6. If the consistency of the edge

is perfect, the max count of edge consistency is the

same as 𝑅

. That is, ECCTC becomes 0. Our final

stereo matching cost (𝑇

) is calculated as the sum

of CTC (𝑇

) and ECCTC (𝑇

) as in (2). The

Figure 6: Example of the equal Census Transform cost for

different patterns.

3D Mapping of Indoor Parking Space Using Edge Consistency Census Transform Stereo Odometry

1017

Figure 7: Example of the equal Census Transform cost for different patterns.

proposed ECCT method has an efficient amount of

computation because it calculates only the edge part

even if the size of the window size increases.

𝑅

𝑛1

∗4 (1)

𝑇

𝑇

𝑇

(2)

4 VO-LIDAR INTEGRATION

The data from both sensors can be integrated using

the coordinate system transformation matrix between

the camera and lidar. We project the 2D lidar data

onto the visual odometry of the same time. This

method can express 3D space using visual odometry

and 2D lidar. Equation 3 is a formula for this. 𝑃

𝑥

,𝑦

,1,1

represents a point in lidar data

measured at time 𝑡 based on the 2D lidar coordinate

system. 𝑇

represents the 4x4 transformation matrix

between the camera sensor and the 2D lidar sensor.

The coordinate system calibration between the two

sensors was mechanically corrected. 𝑇

is the 4x4

transformation matrix of the camera sensor for time t

from the world coordinate system. So 𝑃

𝑥

,𝑦

,𝑧

,1

is world coordinate data for 𝑃

measured from 2D lidar at time 𝑡.

𝑃

𝑇

𝑇

𝑃

(3)

Figure 8: Mobile Mapping System to acquire data of

parking spaces and ramp.

5 EXPERIMENTAL RESULTS

To evaluate the performance of the proposed system,

we acquired the dataset from 5 parking ramps and 3

parking spaces by attaching a 3D mapping system to

the car as shown in Fig. 8. Each sensor of the

proposed 3D scanning device is shown in Table 1.

Our camera acquires 20 fps synchronized 688x650

image size data. 2D lidar acquires synchronized data

at 10 fps. Computer specs are intel-core i7-9700k @

3.60 GHz, 16GB RAM.

Table 1: Sensor information of the proposed 3D scanning

device.

Type Manufacture Model description

Camera FILR GS-U3-41C6C-C Global shutter camera

Lens

KOWA LM4NCL

Focal length (3.5mm)

Angle of view

[HorVer]

(117.7°86.7°)

2D Lidar

SLAMTech RPLIDAR S2 1 channel (360 FOV)

We compare the slopes of three parking spaces

and present quantitative data for performance

comparison between the ECCT stereo matching

method and the KLT stereo matching method. Since

we do not have exact Ground-Truth (GT) slope

information for each parking space, we set two

prerequisites. As the first prerequisite, the angle of

inclination of the parking spaces is assumed to be zero

degrees. As a second prerequisite, we assumed that

the normal vector of the plane fitting for the area

where the scanned results of the two methods

completely overlapped is the normal vector of GT.

We calculate the normal vector for two methods of

plane fitting and calculate the angle error by

calculating the dot product of the GT's normal vector.

Table 2 shows the angular error between the normal

vector to the ground in GT and the normal vector to

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

1018

Figure 9: The 3D mapping result for the parking ramps.

Figure 10: The 3D mapping result for the parking spaces.

the 3D reconstructed point cloud ground. The

proposed method shows better performance at the

angle error.

Table 2: Angle Error between normal vectors of the ground

plane.

Stereo

matching

Method

Dataset

1

Dataset

2

Dataset

3

Angle

Error

(degree)

KLT 7.33994 12.26250 10.19682

ETTC 1.79957 2.62571 5.33328

Fig. 9 shows the 3D reconstruction results for the

three parking ramp section datasets. The left image of

each dataset result shows the entry direction of the

MMS. The image on the right shows the height ramp

map results for the reconstructed point cloud data. Fig.

9(a) is the result of the parking ramp dataset for the

scenario going down from the ground floor to the first

basement floor. Fig. 9(b) is the result of the parking

ramp data set for the scenario going up from the 1st

floor to the 4th floor. Fig. 9(c) is the result of the

parking ramp data set for the scenario going up from

the 3rd basement floor to the 1st floor above the

ground. Fig. 10 shows the 3D restoration results for a

parking space without a slope.

6 CONCLUSIONS

This paper presents a device consisting of a

synchronized stereo camera and a 2D Lidar sensor

and the 3D reconstruction method for parking ramp

sections and parking spaces in real time. The

proposed 3D reconstruction method integrates data

by projecting 2D lidar data based on stereo-based

visual odometry. Visual odometry is based on VINS-

Fusion. The odometry using VINS-Fusion's KLT-

based stereo matching can cause drift in the vertical

direction to the ground. So, we proposed and

integrated the Edge Consistency Census Transform

stereo matching method. Edge Consistency Census

Transform is designed to be robust against repeated

3D Mapping of Indoor Parking Space Using Edge Consistency Census Transform Stereo Odometry

1019

patterns by extending the traditional Census

Transform and adding constraints on the consistency

of the Census block edge. The proposed method

scanned the parking ramp section and parking space

and presented qualitative and quantitative results. In

future research, we plan to improve the performance

of odometry by combining the wheel odometry and

an IMU sensor.

ACKNOWLEDGMENT

This work was supported in part by Basic Science

Research Program through the National Research

Foundation of Korea(NRF) funded by the Ministry of

Education (No. 2021R1A6A1A03043144) and part

by the NRF grant funded by the Korea government

(MSIT)(No. 2021R1A2C2009722).

REFERENCES

Bouguet, J. (1999). Pyramidal implementation of the lucas

kanade feature tracker.

Chirca, M., Chapuis, R., & Lenain, R. (2015). Autonomous

Valet Parking System Architecture. 2015 IEEE 18th

International Conference on Intelligent Transportation

Systems, 2619-2624.

Ding, W., Zhang, L., Chen, J., & Shen, S. (2021).

EPSILON: An Efficient Planning System for

Automated Vehicles in Highly Interactive

Environments. IEEE Transactions on Robotics, 38,

1118-1138.

Elhousni, M., Lyu, Y., Zhang, Z., & Huang, X. (2020).

Automatic Building and Labeling of HD Maps with

Deep Learning. AAAI Conference on Artificial

Intelligence.

International Code Council, (2018). Overview of the

International Building Code, https://codes.iccsafe.org/

content/IBC2018.

Kim, K., Cho, S., & Chung, W. (2021). HD Map Update for

Autonomous Driving With Crowdsourced Data. IEEE

Robotics and Automation Letters, 6, 1895-1901.

Lee, S., Park, Y., & Suh, I.H. (2012). Dependable dense

stereo matching by both two-layer recurrent process

and chaining search. 2012 IEEE/RSJ International

Conference on Intelligent Robots and Systems, 5191-

5196.

Liu, P., & Na, J. (2022). A Generalized Hamming Distance

of Sequence Patterns.

Liu, Y., & Collins, R.T. (2000). A computational model for

repeated pattern perception using frieze and wallpaper

groups. Proceedings IEEE Conference on Computer

Vision and Pattern Recognition. CVPR 2000 (Cat.

No.PR00662), 1, 537-544 vol.1.

Lv, Q., Lin, H., Wang, G., Wei, H., & Wang, Y. (2017).

ORB-SLAM-based tracing and 3D reconstruction for

robot using Kinect 2.0. 2017 29th Chinese Control And

Decision Conference (CCDC), 3319-3324.

Qin, T., Chen, T., Chen, Y., & Su, Q. (2020). AVP-SLAM:

Semantic Visual Mapping and Localization for

Autonomous Vehicles in the Parking Lot. 2020

IEEE/RSJ International Conference on Intelligent

Robots and Systems (IROS), 5939-5945.

Qin, T., Pan, J., Cao, S., & Shen, S. (2019). A General

Optimization-based Framework for Local Odometry

Estimation with Multiple Sensors. ArXiv,

abs/1901.03638.

Roh, H.C., Jeong, J., Cho, Y., & Kim, A. (2016). Accurate

Mobile Urban Mapping via Digital Map-Based SLAM.

Sensors (Basel, Switzerland), 16.

Rosten, E., Porter, R.B., & Drummond, T. (2008). Faster

and Better: A Machine Learning Approach to Corner

Detection. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 32, 105-119.

Zabih, R., & Woodfill, J.I. (1994). Non-parametric Local

Transforms for Computing Visual Correspondence.

European Conference on Computer Vision.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

1020