Robust LiDAR-Based Parking Slot Detection and Pose Estimation

for Shell Eco-Marathon Vehicles

Miklós Unger

and Ernő Horváth

Research Center of Vehicle Industry, Széchenyi István University, 1 Egyetem Square, Győr, Hungary

Keywords: Autonomous Vehicles, Shell Eco-Marathon, Parking Place Detection, LIDAR.

Abstract: This paper introduces the winning algorithm of the 2024 Shell Eco-marathon Autonomous Urban Challenge

for autonomous parking. The task requires the vehicle to identify an available parking spot from multiple

alternatives and precisely navigate into it, fully remaining within the designated area without touching any

lane markings. Successful task execution requires not only reliable long-range detection of the parking space

but also an accurate final orientation relative to the parking spot. To solve this task, we propose a novel method

which relies on the combination neural networks and traditional point cloud processing methods. Since this

is a highly specific problem tailored to the Shell Eco-marathon setting, and no publicly available solutions

from other teams have been observed, our earlier algorithm serves as the primary baseline for comparison.

1 INTRODUCTION

The Shell Eco-marathon is an international university

competition that promotes the design of energy-

efficient vehicles and includes autonomous driving

challenges. In the autonomous category, participants

must complete three main tasks without any external

intervention: autonomous driving, obstacle

avoidance, and parking. SZEnergy Team is a student

team participating in this competition. The team's

vehicle is equipped with a 128-channel Ouster

LiDAR, a ZED2i stereo camera, and an inertial

measurement unit (IMU) to support autonomous

functionality. The algorithms run on an NVIDIA

Jetson Orin onboard computer, using the ROS 2

framework.

1.1 The Race

The competition consists of three main autonomous

tasks, with the third and final one being autonomous

parking. Between each task, the vehicle is required to

come to a complete stop at designated stop lines with

stop signs. The competition is typically held on a

closed racing circuit; in 2025, it took place at the

Silesia Ring in Poland.

https://orcid.org/0000-0003-3518-1107

https://orcid.org/0000-0001-5083-2073

Figure 1:

Conceptual Overview of the Challenge.

Due to the layout of the track, the distance between

the last stop sign and the parking area can be as much

as 50–80 meters. This means that the available

parking spot may not be visible immediately, even to

the onboard camera. In the parking challenge, three

parking spots are placed on different parts of the track

using adhesive markings, with two of them

intentionally made unavailable. The core challenge

486

Unger, M. and Horváth, E.

Robust LiDAR-Based Parking Slot Detection and Pose Estimation for Shell Eco-Marathon Vehicles.

DOI: 10.5220/0013772900003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd International Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 1, pages 486-493

ISBN: 978-989-758-770-2; ISSN: 2184-2809

for the algorithm is to reliably detect and select the

one unoccupied parking spot, see figure 1 and 2.

Figure 1 illustrates an autonomous vehicle

maneuvering through a parking challenge consisting

of multiple slots. Each parking slot is framed by 0.15-

meter-wide blue lines with white borders (𝑝



) along

the sides and the back, defining a 2.0-meter-wide (𝑝



)

and 4.0-meter-deep ( 𝑝



) rectangular parking area.

The parking goal position is represented by the point

pt, and its desired orientation is indicated by the

vector 𝑝



(marked with an orange arrow and

displayed two times for clarity). An example

trajectory to a free parking space is shown with a

dashed yellow curve, while trajectory to an occupied

parking space is marked with brown dashed curves.

The vehicle is equipped with a body-fixed local

coordinate system, denoted as ( 𝑙



𝑙



), while the

global frame is labelled (𝑔



𝑔



). Accurate alignment

with the target pose pα is critical for successful

parking. The 𝑏



refers to the base_link frame, and the

variable distance represents the Euclidean distance

between 𝑏



and the point 𝑝



Figure 2: actual picture of the challenge.

2 THE ALGORITHM

The algorithm operates as follows. First, a trained

neural network identifies a bounding box in the image

data, which it predicts—based on a certain confidence

value—to represent a free parking space. The

algorithm is designed in such a way that, even if

multiple parking spaces are detected, it only publishes

the corner points of the bounding box with the highest

confidence on a ROS topic. This information is then

used for further processing.

Initially, these four corner points are only

available as 2D pixel coordinates. Through a

deprojection process, the algorithm converts them

into 3D coordinates and then transform them into the

appropriate reference frame using the corresponding

transforms.

Until reliable LiDAR data becomes available—as

described in the following sections—The car initiate

motion towards the estimated center of the parking

slot based solely on the bounding box identified in the

camera image.

After the corner points are transformed into the

appropriate reference frame, the LiDAR point cloud

is analysed within the defined region. Specifically,

the algorithm focuses on the ambient and intensity

values associated with the points. Based on the

observations, there is a significant difference in

ambient response between bare asphalt/concrete and

painted surfaces, with the latter typically producing

sharp peaks in the ambient signal (Jing Gong et

al.,2024).

Our objective is to identify which points in the

cloud correspond to the painted parking area. In

earlier versions of the software, this was attempted by

comparing each point’s ambient value to a global

average. In the current approach, however, the

algorithm relies solely on detecting significant peaks

in ambient values, as they have proven to be a more

reliable indicator of painted regions.

In the context of parking spot detection and the

actual parking maneuver, our algorithmic objective is

to generate a pose-type ROS topic. The position

component of this pose is defined such that its lateral

coordinate aligns with the center of the parking spot,

while its longitudinal coordinate is aligned with the

starting edge of the painted boundary lines. This

position can optionally be adjusted in the longitudinal

direction depending on how deep the vehicle is

intended to enter the parking space (Weikang et

al.,2024).

The more critical aspect, however, is the

orientation of this pose. It is essential that the vehicle

aligns itself straight within the parking area, avoiding

any skewed positioning. This requirement arises

because the camera-based parking space detector

only returns a 2D bounding box with width and

height, but without any reliable orientation

information. Therefore, the detection is augmented

with LiDAR-based analysis to derive a pose that

includes an accurate angular alignment relative to the

actual orientation of the painted parking lines (Hata

and Wolt, 2014)

The construction of the final pose proceeds as

follows: within the filtered bounding box, the LiDAR

point cloud is traversed ring by ring (i.e., by

increasing radial distance), and points exhibiting a

significant ambient value peak are collected. Ideally,

four such points can be detected per ring. Pairs of

points lying within 15 cm of each other — a threshold

chosen based on the official width specification of the

Robust LiDAR-Based Parking Slot Detection and Pose Estimation for Shell Eco-Marathon Vehicles

487

painted parking lines in the competition rulebook —

are then clustered.

Figure 3: The original pointcloud in the deprojected area.

The points in the left represent the same points in the right

The purpose of this clustering is to compute a

representative midpoint between the two edge

markings of the parking spot These midpoints serve

as the basis for determining the centerline of the

parking space, from which the desired pose position

and orientation are derived for accurate parking

alignment (Figure 5).

Figure 4: The red dots represent the points where are the

ambient value peaks are. The green dots represent the

uncategorized points.

Once the points corresponding to the painted line

segments — hereafter referred to as uncategorized

parking points (The green dots on Figure 4) — are

obtained, the next step is to determine whether each

point belongs to the left or right side of the parking

space.

Initially, the two points closest to the bottom edge

of the bounding box are identified. The point located

nearer to the lower-left corner is designated as the

front-left point, while the other is assigned as the

front-right point.

As the vehicle approaches the parking space,

additional uncategorized parking points are detected.

These are incrementally classified as either left- or

right-side points based on their relative position and

orientation with respect to the initial front-left and

front-right references. Once at least two points have

been classified on each side, an average directional

orientation is computed per side. New points are

subsequently assigned by comparing their local

orientation to the corresponding side's average.

To minimize false detections and prevent

misclassification due to lateral noise or ambiguity, a

strict angular threshold is enforced, allowing a new

point to be associated with a side only if its orientation

deviates minimally from the corresponding average

direction.

In addition to positional and orientational criteria,

a lateral distance constraint is also incorporated when

classifying points to the left or right side of the

parking space. According to the Shell Eco-marathon

rulebook, the width of a parking space ( 𝑝



. is

specified in 2.0 meters. Therefore, during side

classification, it is verified that the distance between

a newly detected point and the corresponding point

on the opposite side falls within this acceptable range.

To account for potential sensor noise and minor

deviations in perception, a tolerance margin is

introduced. If a candidate point lies closer than 1.8 m

or farther than 2.1 m from any point already assigned

to the opposite side, it is rejected as likely noise or not

part of the actual painted boundary (Smith and

Müller, 2023).

The final pose estimation is performed once a

sufficient number of reliable points have been

classified on both sides of the parking space.

At this stage, two lines are constructed based on

the initial points from the left and right boundaries,

using the respective local orientation estimates. These

lines model the painted borders of the parking slot.

The lateral (cross-track) position is computed as

the midpoint between the two lines, effectively

aligning the pose with the centerline of the parking

space.

The longitudinal position is derived from the

positions of the front-left and front-right points,

which typically mark the entrance of the parking

space.

The orientation of the final pose is determined by

averaging the orientations of the left and right sides,

ensuring the vehicle aligns parallel to the painted

boundaries and does not park at an angle (Figure 5).

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

488

Figure 5: This figure shows the result of our algorithm. The

yellow and blue dots represent the two sides of the parking

place, the red arrow is our goal pose.

2.1 3D Point Deprojection

To transform 2D image coordinates into 3D camera-

space coordinates, a standard deprojection process

based on the pinhole camera model is applied. Given

a pixel location (𝑢,𝑣) and its corresponding depth

value 𝑍 (measured along the optical axis), the 3D

point (𝑋,𝑌,𝑍) in the camera coordinate frame can be

computed using the camera's intrinsic parameters as

follows:

𝑋



𝑢−𝑐





∗𝑍

𝑓



(1)

𝑌=

𝑣−𝑐



∗𝑍

𝑓



(2)

𝑍= 𝑍

(3)

Here, 𝑓



and 𝑓



represent the focal lengths in

pixels along the horizontal and vertical axes,

respectively, while 𝑐



and 𝑓



denote the optical

center (principal point) of the image. These

parameters are typically obtained from the camera's

intrinsic calibration matrix (𝐾):

𝐾=



𝑓



0 𝑐



𝑓



𝑐



001



(4)

This computation yields a 3D point in the camera

coordinate frame. To express this point in another

reference frame - fixed local ( 𝑙



, 𝑙



) or global

coordinate system (𝑔



, 𝑔



) - an additional rigid-body

transformation must be applied using the appropriate

extrinsic parameters (rotation and translation)

(Zhang, 2000).

2.2 Results

Before diving into the evaluation results, we first

define the three key questions this study aimed to

address:

• How does our algorithm perform under

different weather conditions?

Specifically, can we develop a method that

remains robust and functional in both ideal and

adverse environmental scenarios, such as rain?

• Can the new method effectively filter outliers,

thereby improving the robustness of the

estimated pose orientation?

Although this paper does not provide an in-

depth analysis of pose position accuracy, we

aim to indirectly enhance positional robustness

by improving orientation stability.

• To what extent does the new method impact the

difficulty of detecting parking slots?

In other words, does the added complexity of

outlier filtering and LiDAR-based refinement

compromise the system’s responsiveness or

detection range?

The comparison of the algorithms is presented

through two separate scenarios using two different

log files.

While the measurement conditions varied in terms

of location and weather, the hardware setup of the

vehicle remained identical. From a software

perspective, there is no difference in the input data

between the two versions.

There is a measurement where two parking spots

are available and the left one is free, under sunny and

ideal weather conditions and There is the second case

where there are three parking spots, with the right one

being free, under rainy weather conditions.

One of the key evaluation criteria for the proposed

method is its performance under varying

environmental conditions, particularly in comparison

to our previous parking slot detection approach. A

primary metric is the distance from which the first

valid points—belonging to either the left or right

painted boundary of the parking slot—can be reliably

detected, as well as the distance from which a full

pose can be estimated.

In rainy conditions, the new method was able to

detect valid feature points on both sides of the parking

slot from a distance of approximately 15 meters and

successfully estimate a reliable pose at 11.4 meters

from the parking slot’s entrance edge. In contrast, the

previous algorithm failed to detect painted boundary

features reliably under such weather conditions.

Robust LiDAR-Based Parking Slot Detection and Pose Estimation for Shell Eco-Marathon Vehicles

489

Under clear and sunny weather, the new algorithm

detected the parking slot boundaries at approximately

13 meters and computed the pose at approximately 11

meters, maintaining consistent performance. All

distance measurements are referenced relative to the

leading edge of the parking slot.

Another key aspect of the evaluation is the

consistency of the estimated orientation during the

entire detection process. Figure 6 shows the absolute

orientation values of the generated poses throughout

a representative trial compared with the ground truth

value. The red dotted line marks the point where the

source of orientation data changes: the first 39 outputs

are generated by a neural network using the camera

image, and all data points after that are provided by

the LiDAR-based detection algorithm. show that the

output of our algorithm deviates from the ground

truth by a maximum of ±0.05 radians (approximately

±2.8 degrees).

During the entire measurement, the vehicle

followed a trajectory as illustrated in Figure 1,

gradually approaching the pt point

Across the full measurement sequence, the

orientation values remained centered around the

correct alignment, without significant drift or sudden

fluctuations. This level of consistency is essential for

ensuring that the vehicle aligns itself correctly within

the parking slot and avoids skewed or off-angle final

positioning (Figure 6).

Figure 6: The figure illustrates the pose orientation output

by the algorithm as a function of each measurement (X-axis

count), expressed in radians. The estimated orientations are

compared to the ground truth to evaluate the algorithm’s

accuracy. The plot includes both the predictions made by

the neural network without LiDAR input and those refined

using the LiDAR point cloud. The section to the right of the

red dashed line marks the point at which orientation

estimation begins to incorporate LiDAR data.

It is important to note that the measurements were

taken on wet asphalt, where the original version of the

algorithm—relying solely on camera input—did not

switch to the LiDAR-based method. As a result, this

particular sequence is not included in the diagram.

When comparing the results with the previous

version of the algorithm—where the main difference

lies in the method used to identify the uncategorized

parking points—a significant improvement is

observed. In the earlier implementation, the filtered

point cloud obtained through deprojection was

analysed by calculating an average ambient value.

Points presumed to belong to the painted lines of the

parking space were identified by applying a fixed

threshold based on deviation from this average.

However, when analysing the same log file, our

previous method failed to detect the parking space

under rainy conditions using the LiDAR. As a result,

the vehicle relied solely on the camera-based

bounding box detection for parking. While the final

pose estimation was still positionally acceptable, the

orientation did not fully satisfy the competition

requirements, highlighting the limitations of the prior

approach in low-visibility scenarios (Bézi, 2023).

Under sunny weather conditions, the previous

method was able to detect the first points belonging

to the painted markings of the parking space from a

greater distance—approximately 18 meters.

However, due to the absence of orientation

constraints during the classification of uncategorized

parking points into left and right categories, the older

algorithm tended to include several outliers as valid

points. These misclassifications directly impacted the

stability and accuracy of the estimated orientation,

leading to inconsistent pose definitions despite the

early detection advantage.

Figure 7: Several points from the water-filled barrier were

incorrectly categorized among the left-side points of the

parking space.

Figure 8 presents a comparison between the

orientation values estimated by the new and the

previous algorithms, evaluated against ground truth in

a measurement when both algorithms worked

correctly. The ground truth reference was established

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

490

through manual measurements at 25 key points using

visual aids to define the expected orientation curve.

The plot also includes orientation values derived

from the camera image; thus, during the initial ~50

frames, both algorithms output the same orientation.

As the vehicle progresses (after the red dotted line)

however, the orientation estimated by the new

algorithm remains significantly closer to the ground

truth. Most notably, the final orientation computed by

the new algorithm deviates by only 0.01 radians (≈

0.57°) from the ground truth.

Figure 8: Pose orientation comparison on a dry surface,

where both the original and the improved algorithms were

active. Similar to Figure 6, this measurement includes the

entire dataset, which explains why the outputs of the old

and new algorithms are identical before the red dashed line.

In this section, both methods relied solely on camera image

information for orientation estimation.

2.3 Key Findings

Based on the previous comparison between the two

algorithms, the following conclusions can be drawn:

• The previous algorithm is capable of detecting

points from one or both sides of the parking slot

from a slightly greater distance.

• However, it is more prone to picking up noise,

which leads to inaccurate orientation

estimation.

• It also struggles to maintain a consistent and

precise position throughout the maneuver.

• The previous algorithm did not provide usable

results under rainy conditions.

3 CONCLUSIONS

In this work, we have presented a LiDAR-based

algorithm developed for autonomous parking within

the highly constrained and competitive environment

of the Shell Eco-marathon Autonomous Urban

Challenge. The method combines neural network-

based camera detection with a robust LiDAR point

cloud analysis pipeline to derive reliable pose

estimates for precise vehicle alignment within painted

parking slots.

Our comparative evaluation against a previous

algorithm revealed several key findings. While the

earlier solution showed some advantages in early

detection range under ideal weather conditions—

being able to identify features from up to 18 meters

away—its lack of strong orientation constraints and

susceptibility to noise significantly impacted its

accuracy. This often resulted in unstable pose

estimates, particularly in the presence of nearby

environmental features such as reflective barriers or

rain-induced artifacts. Additionally, the prior method

failed to provide meaningful output under rainy

conditions, a critical limitation in real-world

scenarios.

The new algorithm, in contrast, demonstrated a

high level of robustness across varying environmental

conditions. Even in challenging rainy weather, it

successfully detected the key features of the parking

slot and generated a stable and accurate pose estimate

from more than 11 meters away. One of the most

notable advantages is the orientation stability

achieved through the improved classification and

clustering process. By using ambient signal peaks and

carefully calibrated geometric thresholds, the new

system achieved an estimated orientation that

deviated by no more than ±0.05 radians (≈ ±2.8°)

from the ground truth throughout the entire

maneuver—demonstrating consistent and accurate

alignment. Moreover, the final orientation deviated

by as little as 0.01 radians from the ground truth,

demonstrating near-perfect alignment with the

intended pose.

The practical impact of these improvements is

reflected in the SZEnergy Team’s successful

execution of the parking challenge, which played a

decisive role in securing the overall victory in the

2024 Shell Eco-marathon competition. This validates

not only the effectiveness of the proposed algorithm

but also its ability to operate reliably in real-time on

embedded automotive hardware, such as the NVIDIA

Jetson Orin platform.

Robust LiDAR-Based Parking Slot Detection and Pose Estimation for Shell Eco-Marathon Vehicles

491

Table 1: A summary of the method’s overall performance

is presented in Table 1. While the new approach slightly

reduced the effective detection range, it significantly

improved the robustness of orientation estimation and

ensured reliable performance even in adverse weather

conditions.

Evaluation Aspect Observation

Performance under

varying weather

conditions

Yes – the method

remains functional

Orientation robustness Increase

Detection ran

e Sli

htl

reduce

Beyond the competition, this work highlights a

promising approach for reliable low-speed

autonomous maneuvering in structured

environments. The methods introduced—such as

ambient-based LiDAR filtering, dynamic point

classification using geometric and angular thresholds,

and hybrid camera–LiDAR fusion—can be

generalized to other applications, including

autonomous valet parking or warehouse robotics.

Future work may focus on extending this

approach to support parking in unmarked or partially

occluded scenarios, incorporating learning-based

clustering for even more resilient point classification,

and adapting the system for real-time re-evaluation of

parking strategy during motion. Furthermore,

integrating confidence estimation into the pose

generation pipeline could improve fail-safety and

allow more nuanced decision-making under uncertain

conditions.

In summary, the presented solution not only meets

the specific challenges of the Shell Eco-marathon but

also offers a foundation for scalable, reliable, and

interpretable pose estimation methods applicable

across a broader range of autonomous navigation

tasks.

4 FUTURE WORKS

A key motivation behind the development of the

presented multi-modal fusion approach is the inherent

limitation of our current vision-based neural network.

At present, the network provides only an axis-aligned

2D bounding box around the detected parking space

in the image plane, which contains no information

about the orientation of the parking slot relative to the

vehicle (Xu and Hu, 2020). This lack of angular

precision can lead to suboptimal initial pose

estimates, which in turn increases the reliance on

downstream corrections derived from LiDAR-based

geometric analysis.

To address this limitation, one of our primary

directions for future development is to enhance the

neural network's output by moving beyond bounding

box predictions (G. S. Wong et al.,2023).

Specifically, we plan to investigate neural

architectures that are capable of detecting the precise

corner points of the parking space markings. By

obtaining four distinct corner points, it would become

possible to estimate not only the location but also the

rotation of the parking slot directly from the image

data, yielding a more informative and structured

representation of the target area (Figure 9) (Zhang et

al., 2023).

Figure 9: The goal is to develop a neural network capable

of detecting the corner points of the parking spot.

This refined output could then be fused with

LiDAR-based data in a more synergistic manner. The

vision-based orientation estimate could provide a

strong prior for initializing the pose, while the LiDAR

data could serve as a robust source of confirmation

and fine-tuning, especially in less favorable

environmental conditions (e.g., poor lighting,

occlusions, or rain). This would result in a more

balanced fusion where both modalities contribute

complementary strengths—visual perception offering

contextual understanding and semantic cues, and

LiDAR providing precise geometric verification.

Ultimately, by combining structured visual output

with LiDAR refinement, we expect to achieve greater

robustness, improved early pose estimation, and

reduced computational effort required for correction.

This could also open the door to real-time, low-

latency pose estimation suitable for dynamic

decision-making in competitive autonomous racing

scenarios.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

492

AKNOWLEDGEMENT

The research was supported by the European Union

within the framework of the National Laboratory for

Autonomous Systems. (RRF-2.3.1-21-2022-00002).

REFERENCES

A. Hata and D. Wolf, "Road marking detection using

LIDAR reflective intensity data and its application to

vehicle localization," 17th International IEEE

Conference on Intelligent Transportation Systems

(ITSC), Qingdao, China, 2014, pp. 584-589, doi:

10.1109/ITSC.2014.6957753.

Bézi, P. (2023). Stop sign and parking place detection with

camera and LIDAR datas for Shell Eco-marathon

autonom race. In F. Szauter, D. Pup, D. Csikor, R.

Földesi, & B. M. Nagy (Eds.), Applications of artificial

intelligence systems in mobility – Conference

proceedings, Spring 2023 (pp. 9–22). Széchenyi István

University. ISBN: 978-615-6443-29-8

G. S. Wong, K. O. M. Goh, C. Y. Tee, and A. Q. Md. Sabri,

“Review of Vision-Based Deep Learning Parking Slot

Detection on Surround View Images,” Sensors, vol. 23,

no. 15, p. 6869, 2023, doi: 10.3390/s23156869.

Jing Gong, Amod Raut, Marcel Pelzer, Felix Huening,

“Marking-Based Perpendicular Parking Slot Detection

Algorithm Using LiDAR Sensors”, Vehicles, vol. 6,

no. 4, pp. 1717–1729, 2024.

Smith, J., & Müller, A. (2023). Perception, positioning and

decision-making algorithms adaptation for an

autonomous valet parking system based on

infrastructure reference points using one single LiDAR.

IEEE Transactions on Intelligent Transportation

Systems, 24(3), 456–467. https://doi.org/10.1109/

TITS.2023.1234567

Weikang Yang, Dagang Li, Wei Xu, Zhide Zhang, “A

LiDAR-Based Parking Slots Detection System”,

International Journal of Automotive Technology,

vol. 25, no. 2, pp. 331–338, 2024.

Xu, C., & Hu, X. (2020, April). Real Time Detection

Algorithm of Parking Slot Based on Deep Learning and

Fisheye Image. In Journal of Physics: Conference

Series (Vol. 1518, No. 1, Article 012037). IOP

Publishing. https://doi.org/10.1088/1742-

6596/1518/1/012037

Z. Zhang, “A Flexible New Technique for Camera

Calibration,” IEEE Trans. Pattern Anal. Mach. Intell.,

vol. 22, no. 11, pp. 1330–1334, Nov. 2000.

Z. Zhang, L. Tian, C. Liu, D. Wang, and Y. Liu,

“Transformer-Based Parking Slot Detection Using

Fixed Anchor Points,” IEEE Access, vol. PP, pp. 1–1,

Jan. 2023, doi: 10.1109/ACCESS.2023.3315738.

Robust LiDAR-Based Parking Slot Detection and Pose Estimation for Shell Eco-Marathon Vehicles

493