International Journal of

Electrical and Electronics Research (IJEER)

Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

An Improved UFLD-V2 Lane Line Recognition Method

LiKang Bo1,2*, Fei Lu Siaw1 , Tzer Hwai Gilbert Thio1
1
Centre for Sustainability in Advanced Electrical and Electronic Systems (CSAEES), Faculty of Engineering, Built Environment
and Information Technology, SEGi University, 47810 Petaling Jaya, Selangor, Malaysia;
2
Hebei Vocational University of Technology and Engineering，No.473, Quannan West Street, Xindu District, Xingtai, Hebei,
China, 054000

*Correspondence: LiKang Bo; [email protected]

ABSTRACT- Lane line recognition remains a crucial component of autonomous driving, particularly under complex
scenarios involving illumination changes and occlusions. This paper presents a structurally efficient and robust improvement of the
UFLD-V2 architecture, designed for real-time and reliable lane detection. The proposed method integrates three lightweight yet
complementary components: (1) Res2Net, replacing the original ResNet backbone, enhances multi-scale feature extraction and
inference efficiency through reparameterization; (2) an Efficient Multi-scale Attention (EMA) module captures fine-grained
contextual details across varying scene complexities; and (3) the Simple Attention Module (SimAM) is applied in the segmentation
head to suppress background noise and improve localization accuracy. Unlike prior work that uses these modules in isolation, we
propose a tailored integration strategy that achieves a favorable trade-off between accuracy and computational cost. Extensive
experiments on the TuSimple dataset show the effectiveness of our method, achieving 0.95947 accuracy, 0.0262 false positive rate,
0.02328 false negative rate, and an F1 score of 0.96517. Our approach surpasses several state-of-the-art models, including UFLD,
PolyLaneNet, EL-GAN, SAD, CurveFormer++, BEVLaneDet, and PersFormer, particularly under challenging conditions. These
findings highlight the potential of our approach for practical deployment in intelligent driving systems.

Keywords: Lane Line Recognition, UFLD-V2, Res2Net, EMA Attention Module, SimAM Module.

localization. However, despite its solid baseline performance,

ARTICLE INFORMATION UFLD-V2 still exhibits several limitations:
Author(s): LiKang Bo, Fei Lu Siaw, Tzer Hwai Gilbert Thio; • Insufficient multi-scale feature representation, which hinders
Received: 21/01/2025; Accepted: 25/04/2025; Published: 18/06/2025; its adaptability to varying lane shapes and scales;
E- ISSN: 2347-470X;
Paper Id: IJEER 2101-05;
• Limited attention modeling, which weakens its ability to focus
Citation: 10.37391/ijeer.130211 on important spatial and semantic information;
Webpage-link: • Inadequate robustness in the presence of low visibility,
https://ijeer.forexjournal.co.in/archive/volume-13/ijeer-130211.html background noise, or occluded lanes.
Publisher’s Note: FOREX Publication stays neutral with regard to To address these challenges, we propose an enhanced UFLD-
jurisdictional claims in Published maps and institutional affiliations.
V2-based framework that integrates three lightweight yet
complementary modules:
░ 1. INTRODUCTION 1. A Res2Net backbone to improve multi-scale feature
With the rapid development of autonomous driving technology, extraction and hierarchical representation;
lane line detection has become a fundamental task for ensuring 2. An Efficient Multi-Scale Attention (EMA) module to refine
vehicle safety and enabling accurate path planning [1]. In real- context modeling across spatial and channel dimensions;
world environments, however, lane detection remains highly 3. A Simple Attention Module (SimAM) in the segmentation
susceptible to performance degradation under challenging head to suppress background interference and enhance
conditions such as illumination variation, occlusion, and road localization accuracy.
surface clutter [2]. Traditional rule-based or handcrafted-feature
Unlike previous works that employ such modules in isolation,
methods typically lack the adaptability and robustness required
our design implements a task-specific, latency-aware
for such scenarios.
integration strategy. We also apply reparameterization
Recent advances in deep learning, especially those involving techniques to the backbone network during inference to
convolutional neural networks (CNNs), have led to significant compress multi-branch paths into a single branch, thereby
improvements in lane detection performance. Among these, preserving real-time performance.
UFLD-V2 has emerged as a highly regarded framework due to
This design achieves a favorable balance between detection
its balance between structural efficiency and detection
accuracy, model robustness, and computational efficiency,
accuracy. Unlike earlier approaches such as UFLD [3], UFLD-
making it well-suited for practical deployment in intelligent
V2[4] introduces a dedicated detection head and adaptive
driving systems.
anchor mechanisms that enable fast, structured lane

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 277

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

The rest of the paper is organized as follows: Section 2 reviews robustness against illumination changes, occlusions, and road
related work, section 3 introduces the proposed method, section texture variations. End-to-end frameworks such as instance
4 presents experimental results, and Section 5 concludes the segmentation models [15, 16], generative adversarial methods
paper. like EL-GAN[16], and regression-based networks like
PolyLaneNet[17] have achieved strong performance on
░ 2. RELATED WORKS benchmarks like TuSimple and CULane.
Lane detection, as a crucial component of autonomous driving
More recently, a number of UFLD-based extensions have been
systems, plays a vital role in ensuring vehicle safety, supporting
proposed to further improve accuracy and robustness while
decision-making, and maintaining driving stability[5].
maintaining real-time efficiency. MCA-UFLD[18], for
Consequently, it has become a research hotspot in the fields of
example, utilizes a lightweight MobileNetV2 backbone with
intelligent transportation and computer vision in recent years.
coordinate attention and a vanishing point branch to enhance
Traditional lane detection methods primarily rely on
semantic perception without sacrificing speed. Other methods
handcrafted features and classical image processing techniques,
incorporate Split-Attention and ASFF modules to enhance
such as color thresholding, edge detection, and the Hough
multi-scale feature fusion, resulting in notable gains in F1-score
transform. These approaches can deliver acceptable
on CULane[8]. Additionally, techniques such as random
performance under controlled conditions; however, they often
masking and smooth curve loss have been employed to improve
lack robustness in complex environments involving shadow
performance under occlusions and enhance the continuity of
interference, illumination changes, lane occlusions, and road
predicted lane lines. To further increase generalization under
texture variability. This limits their adaptability to diverse and
poor visibility, contrastive learning strategies such as CLLD
unstructured road scenarios. With the rapid development of
have also been introduced[19].
deep learning, convolutional neural network (CNN)-based
methods have become dominant[6]. These methods enable end- These studies highlight the effectiveness of integrating
to-end learning, automatically extracting multi-scale features lightweight backbones, attention mechanisms, and auxiliary
from input images—ranging from low-level textures to high- learning strategies for lane detection. Inspired by these
level structural semantics—leading to more accurate and stable advances, we propose a structurally enhanced UFLD-V2
lane detection under challenging conditions[7]. Furthermore, framework that incorporates a Res2Net backbone for multi-
the integration of attention mechanisms and self-supervised scale feature extraction, an Efficient Multi-scale Attention
learning techniques in recent work has further improved (EMA) module for refined spatial-channel modeling, and a
generalization and robustness, enabling more reliable Simple Attention Module (SimAM) to suppress background
performance in real-world driving scenes[8]. interference. Unlike prior works that use such modules
independently, our method adopts a task-driven, latency-aware
Feature-based Lane detection methods primarily rely on
integration strategy designed to maintain real-time applicability
handcrafted features and classical image processing pipelines.
while significantly improving robustness and accuracy.
For example, some approaches utilize color thresholding in
Detailed descriptions are provided in section 3.
HSV or CIELab color spaces, edge detection, and the Hough
transform to extract lane markings[9-11]. These techniques can
achieve good performance under stable lighting and clear road ░ 3. PROPOSED METHODOLOGY
conditions. However, their strong dependence on color and
edge contrast limits their adaptability in real-world
environments with variable illumination, shadowing, or blurred
markings.

Model-based methods incorporate geometric constraints to

represent lane trajectories using straight lines, parabolas, or
splines. Lane boundaries are typically extracted using
algorithms such as RANSAC, least squares fitting, or vanishing
point estimation[12-14]. These methods provide interpretable
results and can perform well in structured road scenarios.
Nevertheless, they often assume ideal road geometries or
consistent lane visibility, which makes them less effective in
unstructured or highly dynamic environments.

Deep learning-based lane detection methods have become

dominant due to their ability to learn hierarchical visual features
and adapt to complex driving environments. Compared with Figure 1. UFLDv2 network architecture
traditional techniques, CNN-based approaches offer greater

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 278

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

Figure 1 is the basic network studied in this paper. It consists of 3.2. Introducing EMA Attention Module
two main branches: existence branch and localization branch.
The input image is first extracted through the backbone network
to extract deep features, and then a multi-layer perceptron
(MLP) is used to determine whether there is a lane in the image.
If there is a lane, the localization branch is responsible for
predicting the specific location of the lane. Finally, the final lane
location is output.

To enhance the robustness and precision of intelligent vehicle

lane line detection in complex scenarios, this paper optimizes
the UFLD-V2 model, proposing an innovative method that
combines multi-scale feature fusion and attention mechanisms.
The method fully utilizes feature extraction network capabilities
and efficient attention mechanisms, optimizing and
reconstructing the original model's backbone network, feature
extraction modules, and segmentation head. Specific
improvements are as follows:

3.1. Optimizing Backbone Feature Extraction

Network

Figure 3. EMA module

The Efficient Multi-Scale Attention (EMA) module enhances

lane detection under complex scenarios by fusing multi-scale
spatial context and reweighting feature channels dynamically.
It consists of multi-scale convolutional paths and channel-wise
attention computation.

First, the input feature 𝐹 ∈ ℛ 𝐶×𝐻×𝑊 is passed through multiple

convolution branches:

𝐹1 = 𝐶𝑜𝑛𝑣1×1 (𝐹), 𝐹3 = 𝐶𝑜𝑛𝑣3×3 (𝐹), 𝐹5 = 𝐶𝑜𝑛𝑣5×5 (𝐹) (1)

These are concatenated and fused:

Figure 2. Bottleneck block and Res2Net module
𝐹𝑓𝑢𝑠𝑒𝑑 = 𝐶𝑜𝑛𝑐𝑎𝑡(𝐹1 , 𝐹3 , 𝐹5 ) ⟶ 𝐵𝑁 ⟶ 𝑅𝑒𝐿𝑈 (2)
In the UFLD-V2 model, the backbone network has limitations Then, a global average pooling generates channel statistics zzz,
in multi-scale feature extraction capabilities. To address this and a two-layer MLP computes attention weights:
limitation, this paper adopts Res2Net[20] to replace ResNet.
Res2Net introduces hierarchical representation of multi-scale 1
∑𝐻 𝑊
𝑧=
𝐻∙𝑊 𝑖=1 ∑𝑗=1 𝐹𝑓𝑢𝑠𝑒𝑑 (𝑖, 𝑗) , 𝑎 = 𝜎(𝑊2 ∙ 𝑅𝑒𝐿𝑈(𝑊1 ∙ 𝑧))
features, efficiently capturing local and global information in
(3)
complex scenarios. Compared to traditional ResNet, Res2Net
enhances feature expression capabilities through multi-scale Finally, the output is obtained by applying the learned weights:
group convolutions, better adapting to complex road
environments. Additionally, this paper uses reparameterization 𝐹𝑜𝑢𝑡 = 𝐹𝑓𝑢𝑠𝑒𝑑 ⨀𝑎 (4)
technology to optimize Res2Net. During the training phase,
multi-branch structures enhance learning capabilities, while in EMA is embedded into the output of the deepest stage of
the inference phase, multiple branches are merged into a single Res2Net in our model, enabling enhanced detail recognition and
branch, significantly improving inference efficiency and robustness in scenarios with occlusion, lighting variation, or
reducing computational overhead. cluttered backgrounds.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 279

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

3.3. Optimizing Segmentation Head with SimAM UFLD-V2 in multiple complex scenarios, verifying the
Module effectiveness and superiority of the proposed method.

3.4. Integration Strategy and Efficiency

Considerations
To comprehensively improve robustness, accuracy, and
efficiency in complex scenarios, this paper integrates three
modules—Res2Net, EMA, and SimAM—into the UFLD-V2
Figure 4. SimAM module
framework. Each module plays a distinct yet complementary
role in addressing the limitations of the original model.
The Similarity Attention Mechanism (SimAM) is a lightweight, Res2Net, replacing the original ResNet backbone, enhances
parameter-free attention module inspired by neuroscience, multi-scale feature extraction by introducing hierarchical
which evaluates the importance of each neuron in a feature map. residual-like connections. During the training phase, its multi-
Unlike traditional attention methods (e.g., SE-Net), SimAM branch structure improves the learning capability. During
does not rely on convolution or MLP operations, thus ensuring inference, reparameterization is applied to merge branches into
low computational cost and real-time performance. a single-path structure, significantly accelerating prediction and
The core idea of SimAM is to define an energy function 𝐸𝑖 for reducing computational cost.
each neuron 𝑥𝑖 , which consists of its deviation from the mean The EMA (Efficient Multi-scale Attention) module is added to
and suppression from other neurons: both low-level and high-level feature maps within the Res2Net
backbone. By dynamically fusing multi-scale spatial context
𝐸𝑖 = (𝑥𝑖 − 𝜇)2 + 𝜆 ∑𝑗≠𝑖 𝑥𝑗2 (5)
and computing adaptive channel attention, EMA effectively
enhances fine-grained feature representation in scenarios
where 𝜇 is the average response of all neurons in the same involving occlusion, lighting variations, and background
channel and 𝜆 is a suppression coefficient. The importance clutter.
score is computed as:
SimAM, a parameter-free lightweight attention mechanism, is
1
𝛼𝑖 = (6) inserted after the pooling layer in the segmentation head. By
1+𝑒 −𝐸𝑖
computing importance scores for each neuron based on an
This scalar 𝛼𝑖 is used to reweight the neuron's activation, energy function, SimAM suppresses irrelevant background and
thereby enhancing foreground features and suppressing enhances the localization of lane boundaries without additional
irrelevant background. In our model, SimAM is inserted after computational burden.
the pooling layer in the segmentation head, significantly
This integration strategy results in a structurally efficient and
improving the detection precision for complex scenarios.
inference-friendly model. The proposed method was trained on
the TuSimple dataset, with the input resolution downscaled to
320×800 to improve speed. The training was conducted using
200 epochs, batch size of 16, Adam optimizer with a learning
rate of 0.05 (adjusted using cosine annealing), and a weight
decay of 0.0005. All experiments were performed under Ubuntu
20.04, using the PyTorch framework and NVIDIA H800 GPU.

Notably, the improved model achieves an inference speed of

235 FPS on RTX 2080Ti-equivalent hardware, which is only
slightly lower than UFLD-V2 (260 FPS), while significantly
outperforming other attention or transformer-based models
such as CurveFormer++ (110 FPS). These results confirm that
the improved architecture maintains real-time processing
capability while achieving better robustness and accuracy,
making it well-suited for practical deployment in autonomous
driving systems.
Figure 5. Improved Network Architecture

By optimizing the backbone network, introducing multi-scale ░ 4. RESULTS

attention mechanisms, and combining lightweight attention This section presents the experimental setup, data preprocessing
modules, the lane line detection method proposed in this paper steps, evaluation metrics, ablation study, performance
demonstrates stronger robustness and higher detection precision benchmarking, inference efficiency, and visual analysis of the
when facing complex scenarios. Experimental results show that proposed method. All experiments follow the configuration
the improved model significantly outperforms the original

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 280

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

detailed in section 3.4, ensuring consistency between defined threshold, the image remains unchanged. Otherwise,
architectural design and training procedures. the algorithm:

4.1. Training Settings • Computes the intensity range between the 5th and 95th
The model was trained on the TuSimple dataset, with the percentiles of pixel values;
original image resolution (1280×720) downscaled to 320×800 • Discards pixels outside this range to remove outliers;
to improve computational efficiency. Training was conducted • Applies linear contrast stretching or gamma correction to
for 200 epochs using a batch size of 16. The Adam optimizer enhance important visual details;
was employed with an initial learning rate of 0.05, which was • Outputs the adjusted image for model training.
decayed using a cosine annealing schedule. A weight decay of
This preprocessing strategy selectively enhances dark or
0.0005 was applied to mitigate overfitting. All experiments
shadowed images during training, ensuring lane markings
were conducted on Ubuntu 20.04 using the PyTorch framework
remain visible while maintaining a lightweight and efficient
and an NVIDIA H800 GPU.
training pipeline.
To improve model generalization and reduce overfitting, a The ABCE module enhances underexposed training images by
variety of data augmentation techniques were applied online removing intensity outliers and normalizing contrast, improving
during training via PyTorch’s data pipeline. These include: lane visibility without increasing model complexity.

• Random horizontal flipping, which simulates mirrored

driving scenarios and enhances robustness to diverse road
geometries;
• Random brightness and contrast adjustment, which
introduces lighting variation to help the model learn under
changing illumination, such as shadows, overexposure, and
sunlight reflections;
• Gaussian noise injection and slight affine transformations,
which simulate real-world camera disturbances and
environmental imperfections.

(a)

(b)
Figure 7. Visual comparison before and after applying ABCE
algorithm
Figure 6. Flowchart of the Adaptive Brightness Compensation The left image shows the original low-light input from the
Enhancement (ABCE) algorithm
TuSimple dataset, where lane markings are difficult to
distinguish due to poor illumination. The right image displays
Additionally, we introduce a targeted low-light enhancement
the result after applying the proposed adaptive brightness
method called adaptive brightness compensation enhancement
compensation enhancement (ABCE) algorithm, which
(ABCE) to further improve robustness under poor illumination.
improves local contrast and lane visibility. This preprocessing
As illustrated in figure 6, ABCE first computes the grayscale
enhances the input quality before lane detection, especially
brightness of the input image. If the brightness is above a
under shadowed or underexposed conditions.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 281

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

4.2. Dataset Preprocessing and Label Generation False positive rate: This value is the ratio of non-lane points
To facilitate structured training, the original TuSimple dataset predicted by the model as lane line points.
annotations in JSON format were converted into a standardized 𝐹𝑝𝑟𝑒𝑑
format using a custom preprocessing script 𝐹𝑃 = (9)
𝑁𝑝𝑟𝑒𝑑
(convert_tusimple.py). The process involves the following key
steps: Among them, 𝐹𝑝𝑟𝑒𝑑 represents the number of lane line points
• Lane classification: Lane lines are extracted from the that are misdetected during prediction, and 𝑁𝑝𝑟𝑒𝑑 represents the
annotation files and classified based on their slope. Lanes total number of lane line points in the prediction results.
are sorted from left to right, and lines with a length shorter
than 90 pixels are filtered out to exclude incomplete or F1 score: An indicator that comprehensively evaluates the
irrelevant annotations. performance of a classification model. It is often used to
• Segmentation label generation: For each input image, a measure the balance between the accuracy and recall of binary
corresponding segmentation mask is generated where and multi-classification models. It is the harmonic average of
pixel values {1, 2, 3, 4} represent the four lane lines from the two.
left to right. These masks serve as the ground truth for
segmentation-based training. (10)
• Structured label files: Two text files are generated—
train_gt.txt for training and test.txt for testing. The 𝑇𝑃 𝑇𝑃
Among them, 𝑃 = ,𝑅 = TP represents the
training file records the path to each image and the binary 𝑇𝑃+𝐹𝑃 𝑇𝑃+𝐹𝑁
existence indicators for lane positions, while the test file number of samples correctly recognized by the model, and TN
contains paths to test images for evaluation without represents the number of negative samples recognized by the
model as not lane lines.
labels.
• Annotation caching: Parsed Lane information is saved
4.4. Performance Curves
into a cache file (tusimple_anno_cache.json) to reduce
Figure 8 shows the training progression of accuracy, false
preprocessing time during repeated training runs. positives, false negatives, and F1 score over 200 epochs. The
model demonstrates stable convergence and consistent
This preprocessing step ensures that the lane annotations are
performance gains across all metrics.
consistent, structured, and suitable for deep learning-based lane
detection frameworks, while remaining fully compatible with
the official TuSimple evaluation protocol.

4.3. Evaluation indicators

This paper uses the following evaluation indicators commonly
used in the field of object detection to evaluate the lane line
detection performance: Accuracy (Acc), False Negative (FN),
False Positive (FP) and F1 score. The definitions of each
indicator are as follows:

Accuracy: used to measure the ability of the model to make

correct predictions.

(7)

𝐶𝑐𝑙𝑖𝑝 represents the number of lane line points correctly

predicted by the model, and 𝑆𝑠𝑙𝑖𝑝 represents the number of lane
line points in the corresponding group-truth.

Missed detection rate: This value is the proportion of lane line

points identified as non-lane line points in the model's
prediction.
𝑀𝑝𝑟𝑒𝑑
𝐹𝑁 = (8)
𝑁𝑔𝑡

Among them, 𝑀𝑝𝑟𝑒𝑑 represents the number of lane line points

missed by the model in the prediction, and 𝑁𝑔𝑡 represents the
number of all lane line points in the label.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 282

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

Figure 8. Training curves of Accuracy, False Positive Rate, False Negative Rate, and F1 Score over the first 100 epochs. The complete training
spanned 200 epochs, with continued performance refinement observed beyond the 100-epoch mark. Only the initial portion is visualized here for
clarity

4.5. Ablation experiment

Table 1 presents an ablation study evaluating the individual and combined effects of Res2Net, EMA, and SimAM. Performance is
recorded at each stage, showing incremental improvements

░ Table 1. Comparison of datasets at different stages

Res2Net EMA SimAM Accuracy (%) FP (%) FN (%) F1 Score (%)

– – – 95.79 2.94 4.85 96.09

√ – – 95.66 2.69 4.81 96.23

– √ – 95.83 2.80 4.58 96.20

– – √ 95.81 2.75 4.62 96.18

√ √ – 95.85 2.93 4.52 96.26

√ – √ 95.89 2.65 4.06 96.38

– √ √ 95.87 2.59 3.85 96.41

√ √ √ 95.95↑ 2.62↓ 2.33↓ 96.52↑

(Note: All results are averaged over three repeated runs under fixed random seeds)

Table 1 presents the results of ablation experiments conducted

to evaluate the individual and combined contributions of the The best performance is achieved when all three modules are
Res2Net backbone, the Efficient Multi-scale Attention (EMA) integrated (last row), resulting in an accuracy of 95.95%, FP
module, and the Simple Attention Module (SimAM). The rate of 2.62%, FN rate of 2.33%, and an F1 score of 96.52%.
baseline model (first row) corresponds to the original UFLD-
V2 framework without any modifications. These results confirm the effectiveness of the proposed
integration strategy: the combination of a multi-scale backbone
From the results, it is evident that each module contributes design, attention-based enhancement, and spatial noise
positively to the overall performance. When introduced suppression significantly improves both precision and
individually, all three components lead to improvements in robustness in lane detection.
either accuracy or robustness. Among the single-module
variants, SimAM provides the most significant gains, reducing To further evaluate the computational feasibility of our method,
both the false positive rate (FP) and false negative rate (FN) we report its model size, FLOPs, and inference speed in table 2.
from 2.94% and 4.85% to 2.75% and 4.62%, respectively. Despite a significantly higher parameter count due to the deeper
backbone and attention modules, our model maintains real-time
The integration of Res2Net and EMA shows clear performance (235 FPS), demonstrating its suitability for
complementary benefits—enhancing both multi-scale feature deployment in autonomous driving systems.
extraction and contextual attention modeling.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 283

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

░ Table 2. Model Complexity and Inference Efficiency percentage points, reduces FP and FN rates by up to 0.97% and
Params FLOPs 0.54%, respectively, and enhances F1 score by up to 1.45
Model Backbone FPS Notes percentage points.
(M) (G)
UFLD-V2 ResNet- As reported
(Baseline) 18
10.8 13.6 260
in [3]
These results demonstrate that the proposed method not only
achieves state-of-the-art detection accuracy, but also offers
Ours (Full Res2Net-
Estimated greater robustness and reliability across diverse scenarios,
~200.0 ~35–50 235 from pth making it well-suited for practical deployment in autonomous
Model) 50
and testing driving systems.
(Note: Parameter counts vary with backbone size. UFLD-V2 uses ResNet-
18, while the proposed model adopts Res2Net-50 along with EMA and SimAM 4.7. Inference Speed Evaluation
modules. FLOPs are approximate estimates based on architectural analysis
(input size 320×800). FPS is measured on an NVIDIA RTX 2080Ti using batch To evaluate the real-time performance of our method, we
size = 1). compare the inference speed (FPS) of the proposed Enhanced
UFLD-V2 with several representative lane detection models.
4.6. Benchmark Comparison For fairness, all FPS values are measured or estimated under a
Table 3 compares the improved model against baseline and unified hardware setting equivalent to RTX 2080Ti.
state-of-the-art methods. Our model consistently outperforms
others in all four-evaluation metrics. As shown in table 4, our model achieves an inference speed of
235 FPS, which is slightly lower than the original UFLD-V2
░ Table 3. Comparative performance with existing methods (260 FPS) but significantly higher than other transformer-based
on TuSimple test set methods such as CurveFormer++ (110 FPS). This demonstrates
FN
that the proposed improvements—including Res2Net, EMA,
Accurac FP Rate F1 Score and SimAM—offer substantial accuracy gains without
Model Rate
y (%) (%) (%) sacrificing real-time inference capability. The complete
(%)
UFLD 95.82 19.05 3.92 87.87 comparison is shown below;
PolyLaneNet 93.36 9.42 9.33 90.63
░ Table 4. Inference speed (FPS) comparison of different
EL-GAN 94.90 4.14 3.36 96.26
lane detection models (unified on RTX 2080Ti)
SAD 95.64 60.20 20.50 95.92
Model FPS
CurveFormer++ 95.81 2.75 3.12 96.30
UFLD 326
BEV-LaneDet 95.38 3.56 3.78 95.47
UFLD-V2 260
PersFormer 95.69 3.22 2.99 96.08
PolyLaneNet 115
Improved
95.95↑ 2.62↓ 2.33↓ 96.52↑ CurveFormer++ 110
UFLD-V2
(Note: Metrics for Improved UFLD-V2 are averaged over three runs) Enhanced UFLD-V2 235
(Note: FPS values are either measured directly or scaled from published data
As shown in table 2, the proposed Improved UFLD-V2 based on performance-equivalent conversions to RTX 2080Ti)
achieves the best overall performance across all evaluation
metrics. It obtains the highest accuracy of 95.95%, 4.8. Qualitative Visualization
outperforming the weakest baseline, PolyLaneNet (93.36%), by To assess the robustness and generalization of the proposed
2.59 percentage points. method beyond the TuSimple dataset, we performed qualitative
testing on additional challenging scenes from the CULane and
In terms of false positive rate, our model achieves the lowest BDD100K datasets.
value of 2.62%, which is significantly lower than that of EL-
GAN (4.14%) and dramatically lower than SAD (60.20%), Figure 9 presents the model’s visual outputs in five typical
indicating enhanced capability in suppressing erroneous driving conditions:
detections. (a) Normal: standard daylight from TuSimple
(b) Occlusion: vehicle-blocked lanes from TuSimple
For false negative rate, the proposed model also performs best (c) Hlight and
with only 2.33%, improving upon the next-best EL-GAN
(d) Shadow: strong illumination and contrast from CULane
(3.36%) by 1.03 percentage points and outperforming SAD
(20.50%) by a wide margin. (e) Night: low-light driving scenes from BDD100K

The overall detection quality is reflected in the F1 score, where As observed, the proposed model consistently achieves robust
the Improved UFLD-V2 achieves 96.52%, slightly surpassing and accurate lane detection, maintaining continuity and
EL-GAN (96.26%) and CurveFormer++ (96.30%). reducing false positives even under adverse visual disturbances.
These results provide strong visual evidence of the model’s
Compared with recent transformer-based approaches such as generalization ability across diverse urban and illumination
CurveFormer++, BEV-LaneDet, and PersFormer, our method conditions. Figure 9 Visualization of detection results under
consistently achieves superior performance across all four different conditions: (a) Normal, (b) Occlusion, (c) Road color
metrics. Specifically, it improves accuracy by up to 0.32 variation.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 284

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

(a)
Normal

(b)
Occlusion

(c)
High
Light

(d)
Shadow

(e)
Night

Figure 9. Qualitative lane detection results across five representative scenarios from three datasets: (a) Normal, (b) Occlusion (TuSimple), (c)
High Light (CULane), (d) Shadow (CULane), and (e) Night (BDD100K). Each row displays three sample outputs produced by the proposed
method under the given condition. The results demonstrate the model’s ability to generalize across varying road structures, lighting conditions,
and occlusion patterns.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 285

 International Journal of
Electrical and Electronics Research (IJEER)
Open Access | Rapid and quality publishing Research Article | Volume 13, Issue 2 | Pages 277-286| e-ISSN: 2347-470X

[9] Narote, S.P., et al., A review of recent advances in lane detection and
░ 5. CONCLUSION AND FUTURE departure warning system. Pattern Recognition, 2018. 73: p. 216-234.
WORKS [10] Ma, C.X., et al., A Real-time Semantic Segmentation Model for Lane
This paper presents an enhanced lane detection framework Detection. Journal of Network Intelligence, 2024. 9(4): p. 2234-2257.
based on the UFLD-V2 architecture, incorporating Res2Net for
multi-scale feature extraction, the EMA attention module for [11] Yoo, J.H., et al., A robust lane detection method based on vanishing point
estimation using the relevance of line segments. IEEE Transactions on
robust feature representation in complex scenarios, and the
Intelligent Transportation Systems, 2017. 18(12): p. 3254-3266.
SimAM module for refined lane boundary segmentation.
Experimental results on the TuSimple dataset validate the [12] Wang, J., et al. Lane boundary detection based on parabola model. in
effectiveness of the proposed approach, achieving an accuracy 2010 IEEE International Conference on Information and Automation,
ICIA 2010. 2010.
of 0.95947, a false positive rate of 0.0262, a false negative rate
of 0.02328, and an F1 score of 0.96517, thereby outperforming [13] Wang, J. and X. An. A multi-step curved lane detection algorithm, based
the original UFLD-V2 and several state-of-the-art baselines. on hyperbola-pair model. in 2010 IEEE International Conference on
Automation and Logistics, ICAL 2010. 2010.
The proposed method achieves a favorable balance between
[14] Huang, A.S., et al., Finding multiple lanes in urban road networks with
detection robustness and real-time inference efficiency. Its
vision and lidar. Autonomous Robots, 2009. 26(2-3): p. 103-122.
lightweight design and high FPS performance demonstrate that
the model is well-suited for deployment in practical [15] Neven, D., et al. Towards End-to-End Lane Detection: An Instance
autonomous driving systems, where both accuracy and speed Segmentation Approach. in IEEE Intelligent Vehicles Symposium,
Proceedings. 2018.
are essential for reliable operation in diverse driving conditions.
[16] Ghafoorian, M., et al. EL-GAN: Embedding loss driven generative
Future research will focus on further enhancing the model’s adversarial networks for lane detection. in Lecture Notes in Computer
efficiency, adaptability, and generalization capabilities in real- Science (including subseries Lecture Notes in Artificial Intelligence and
world applications. This includes optimizing inference through Lecture Notes in Bioinformatics). 2019.
network compression and pruning, incorporating domain
[17] Tabelini, L., et al. PolyLaneNet: Lane estimation via deep polynomial
adaptation to improve cross-scenario robustness, integrating regression. in Proceedings - International Conference on Pattern
multimodal sensor inputs (e.g., LiDAR, GPS) for more reliable Recognition. 2020.
environmental perception, exploring advanced attention
[18] Han, L., et al. Lane Detection Method Based on MCA-UFLD. in 2023
mechanisms for spatiotemporal modeling, and leveraging self-
IEEE 8th International Conference on Intelligent Transportation
supervised learning to reduce reliance on large-scale labeled Engineering (ICITE). 2023.
datasets.
[19] Chen, S., & Zhang, Y. (2024, August). Lane detection algorithm based
on improved UFLD. In Fourth International Conference on Image
░ REFERENCES Processing and Intelligent Control (IPIC 2024) (Vol. 13250, pp. 748-
[1] Ni, J., et al., A survey on theories and applications for self-driving cars 755). SPIE.
based on deep learning methods. Applied Sciences (Switzerland), 2020.
10(8). [20] Gao, S.H., et al., Res2Net: A New Multi-Scale Backbone Architecture.
IEEE transactions on pattern analysis and machine intelligence, 2021.
[2] Rui, s. and c. hui, Lane detection algorithm based on geometric matrix 43(2): p. 652-662.
sampling. Chinese Science: Information Science, 2017(47(04)): p. 455-
467. © 2025 by the LiKang Bo, Fei Lu Siaw, Tzer
Hwai Gilbert Thio. Submitted for possible
[3] Qin, Z., H. Wang, and X. Li. Ultra Fast Structure-Aware Deep Lane open access publication under the terms and
Detection. in Lecture Notes in Computer Science (including subseries
conditions of the Creative Commons Attribution (CC BY) license
Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics). 2020. (http://creativecommons.org/licenses/by/4.0/).

[4] Qin, Z., P. Zhang, and X. Li, Ultra Fast Deep Lane Detection With Hybrid
Anchor Driven Ordinal Classification. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2024. 46(5): p. 2555-2568.

[5] Angayarkanni, S.A., R. Sivakumar, and Y.V. Ramana Rao, A review on

traffic congestion detection methodologies and tools. International
Journal of Advanced Science and Technology, 2019. 28(16): p. 1400-
1414.

[6] Sohail, A., et al., Data-driven approaches for road safety: A

comprehensive systematic literature review. Safety Science, 2023. 158.

[7] Mamun, A.A., et al., A Comprehensive Review on Lane Marking

Detection Using Deep Neural Networks. Sensors, 2022. 22(19).

[8] Yang, D., W. Bao, and K. Zheng, Lane Detection of Smart Car based on
Deep Learning. Journal of Physics: Conference Series, 2021. 1873(1): p.
012068.

Website: www.ijeer.forexjournal.co.in An Improved UFLD-V2 Lane Line Recognition Method 286