Using Attention Mechanisms in Compact CNN Models for Improved

Micromobility Safety Through Lane Recognition

Chinmaya Kaundanya

1 a

, Paulo Cesar

2 b

, Barry Cronin

2 c

, Andrew Fleury

2 d

Mingming Liu

1 e

and Suzanne Little

1 f

Insight SFI Centre for Data Analytics, Dublin City University, Ireland

Luna Systems, Dublin, Ireland

ﬂ

{mingming.liu, suzanne.little}@dcu.ie

Keywords:

Micromobility, Lane Recognition, MobileNet, Attention Mechanisms.

Abstract:

The use of personal transportation devices such as e-bikes and e-scooters (micromobility) necessitates the

development of improved safety support systems using highly-accurate, real-time lane recognition. How-

ever, the constrained operating environment, both computationally and physically, on such devices restricts

the applicability of existing sensor-based solutions. One option is to leverage vision-based systems and AI

models. However, these are typically built using high-spec processors and high-memory platforms and the

models need to be adapted to low-spec platforms such as microcontrollers. A signiﬁcant barrier to the de-

velopment and evaluation of these potential solutions is the lack of lane recognition datasets that focus on the

ﬁrst-person (rider) perspective. We contribute a lane recognition dataset of micromobility ﬁrst-person perspec-

tive images from e-mobility rides. This dataset is utilized to assess the impact of channel and spatial attention

on compact CNN models, driven by the aim to maximize utilization through the addition of cost-effective

operations like these attention mechanisms, which introduce only a modest increase in the number of param-

eters. We ﬁnd that adding channel and spatial attention can improve the performance of the standard compact

CNN classiﬁcation models and speciﬁcally that adding the spatial branch improves the performance of the

model with channel attention. The MobileNetV3 model with the fewest parameters among those with channel

plus spatial attention maintained high overall performance. Our code and dataset are publicly accessible at:

https://github.com/Luna-Scooters/Compact-Attention-based-CNNs-on-MLRD.

1 INTRODUCTION

Micromobility is a radical and innovative approach

to minimise the usage of private transportation for

short distances by using personal mobility devices

such as e-bikes and e-scooters. It is a sustainable

alternative to conventional carbon-powered vehicles

while also being ﬂexible and cost-effective. As a re-

sult, it has started gaining traction as a transportation

method worldwide, making it crucial to establish ef-

fective rules and regulations for the use of e-scooters

and e-bikes.

https://orcid.org/0009-0007-4046-5936

https://orcid.org/0009-0000-7171-499X

https://orcid.org/0009-0008-5720-8941

https://orcid.org/0009-0003-6916-6770

https://orcid.org/0000-0002-8988-2104

https://orcid.org/0000-0003-3281-3471

According to a recent report from the Insurance

Institute for Highway Safety (IIHS) (Cicchino et al.,

2021), 60% of e-scooter accidents occur on side-

walks, hence cities are now demanding robust mi-

cromobility safety technology as a minimum licens-

ing requirement. The safe, wide-spread use of these

micromobility solutions requires both regulation and

technological supports such as highly accurate, real-

time, lane detection and localisation. Current GPS

and sensor technologies are unable to provide high

precision or to adapt to variable road structures en-

countered by micromobility vehicles (Fox et al.,

2017). LASER and LiDAR sensors offer potentially

more accurate solutions however are computationally

expensive and difﬁcult to deploy in the highly con-

strained operating environments (Xing et al., 2018).

Artiﬁcial intelligence (AI) based options using im-

age inputs are very promising but still perform best

in controlled environments (i.e., lighting, perspective)

Kaundanya, C., Cesar, P., Cronin, B., Fleury, A., Liu, M. and Little, S.

Using Attention Mechanisms in Compact CNN Models for Improved Micromobility Safety Through Lane Recognition.

DOI: 10.5220/0012600300003702

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

In Proceedings of the 10th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2024), pages 88-98

ISBN: 978-989-758-703-0; ISSN: 2184-495X

and with high-spec computation and memory require-

ments.

Bridging the gap between these challenges and

the needs of micromobility, we turn our focus to

the current state of deep-learning-based lane recog-

nition technologies. The majority of existing deep-

learning-based lane recognition techniques in au-

tonomous vehicles use semantic segmentation (Za-

karia et al., 2023), which does not align well with the

requirements of constrained micromobility environ-

ments. Primarily because segmentation models are

typically large in size, necessitating high memory and

computational power for rapid inference. However,

micromobility vehicles, being more affordable and

having limited physical space and power, offer signif-

icantly less computational capacity, rendering these

models unsuitable.

Given the constraints on resources and to tackle

the speciﬁc challenges of micromobility safety, we

propose a lane recognition strategy that leverages

channel and spatial attention mechanisms for Convo-

lutional Neural Networks (CNN). This lane recogni-

tion approach is aimed at accurately identifying the

lane in which the micromobility rider is traveling in

real-time and subsequently sending necessary alerts.

Attention mechanisms for Convolutional Neural

Networks allow a neural network to focus on rele-

vant input elements and are a vital tool for enhancing

CNN model performance and efﬁciency (Zhu et al.,

2019; Fu et al., 2020). The two primary attention

mechanisms, spatial and channel attention, capture

pixel-level pairwise relationships and channel depen-

dencies respectively (Zhang and Yang, 2021). The

purpose of introducing channel and spatial attention

algorithms is to maximize the performance of com-

pact CNN models, particularly due to the resource-

constrained environment, by utilizing a few additional

trainable parameters with relatively cheap operations.

A major challenge in developing and evaluating

solutions for lane recognition for micromobility sce-

narios is the lack of speciﬁc, labeled datasets with

images from the ﬁrst-person or rider’s perspective.

Previous research focused on urban footpaths utilized

crowd-sourcing to compile a dataset tailored for urban

mobility analysis. This work highlighted the limita-

tions of existing large datasets, which predominantly

feature images of footpaths and bike lanes captured

from vehicle-mounted cameras (GM et al., 2021).

To address these challenges, we have devel-

oped the Micromobility Lane Recognition Dataset

(MLRD), a novel multi-label image classiﬁcation

dataset speciﬁcally designed for the micromobility

safety applications with the ﬁrst person perspective of

the rider. The dataset encodes an information about

the road lane on which rider is riding on and also what

type of road surface, time and weather of the day. The

details of our proposed dataset can be found in sec-

tion 4.

This research is preliminary work exploring the

question: does adding relatively cheap operations

such as channel and spatial attention enable us to im-

prove the performance of compact CNN image clas-

siﬁcation models in constrained environments using

low-resolution input images? Considering these con-

straints, lane classiﬁcation for micromobility is heav-

ily dependent upon learning not only the channel-

wise dependencies but also the spatial relationships

in the feature maps. This rationale led us to speciﬁ-

cally select Squeeze-and-Excitation (Hu et al., 2018)

for channel attention and Coordinate Attention (Hou

et al., 2021) for a more comprehensive mechanism

that encompasses both channel and spatial attention.

In our research, we conducted a series of exper-

iments to assess the impact of integrating attention

mechanisms into compact CNN models for image

classiﬁcation. Notably, the MobileNetV3 (Howard

et al., 2019) model, augmented with channel and spa-

tial attention, demonstrated impressively stable per-

formance metrics. It achieved overall performance

nearly comparable to the baseline model, despite hav-

ing fewer parameters. The MobileNetV2 (Sandler

et al., 2018) channel attention variant showed a slight

improvement in overall precision, while the variant

with both channel and spatial attention exhibited a

signiﬁcant improvement in overall precision and a

slight enhancement in overall performance.

However, there was a noticeable decline in the

recall metric for the “road” class in MobileNetV2

model with channel and spatial attention, and a dete-

rioration in overall performance in the standard Mo-

bileNetV3 model with channel attention. This pat-

tern suggests that while additional parameters from

attention mechanisms can reduce false positives, they

may also lead to potential overﬁtting, impacting the

models’ overall performance. These ﬁndings lead us

to conclude that integrating channel and spatial atten-

tion mechanisms with compact base models does not

straightforwardly enhance their performance. This

highlights the need for more comprehensive research,

especially considering diverse use-cases, to fully un-

derstand and optimize the integration of attention

mechanisms in compact CNN architectures.

The contributions of this paper are as follows:

1. Proposing a real-time lane recognition approach

for micromobility safety applications utilizing a

compact multi-label classiﬁcation model capable

of real-time identiﬁcation of road lane types when

deployed on low-spec microcontrollers.

Using Attention Mechanisms in Compact CNN Models for Improved Micromobility Safety Through Lane Recognition

2. To address the limitations of existing lane recog-

nition datasets, a new multi-label image classiﬁ-

cation dataset with a ﬁrst-person micromobility

rider perspective.

3. Evaluating compact MobileNet models, both with

and without channel and spatial attention.

The remainder of this paper is structured as fol-

lows: Section 2 contains are review of lane recog-

nition approaches using attention mechanisms. The

technical speciﬁcations of the models and algorithms

used are described in Section 3. We describe the de-

tails about the new dataset in Section 4. The exper-

imental methodology and implementation details are

explained in Section 5. The experimental results are

discussed in Section 6. Finally, we draw our conclu-

sions and discuss future work in Section 7.

2 RELATED WORK ON LANE

RECOGNITION

Existing lane recognition efforts have started to utilise

attention mechanisms to improve the detection, seg-

mentation and classiﬁcation of lanes. However they

mostly focus on footage from cars and on detecting

the type and location (relative or absolute) of lanes in

standard road driving scenarios. Here we review the

impact of using attention mechanisms for lane recog-

nition.

Zhang et al. (2021) have developed a real-time

lane recognition system utilizing the Convolutional

Block Attention Module (CBAM) (Woo et al., 2018),

a channel and spatial-based attention mechanism.

Their Convolutional Neural Network architecture is

composed of an encoder designed for lane speciﬁc

feature extraction, one binary decoder and another

decoder to predict the feature maps comprising lane

instances. By integrating CBAM, the encoder ef-

fectively captures intricate details about the targeted

area. This method creates a synergy between the fea-

tures derived from convolution layers and those ac-

quired via attention mechanism, thereby enhancing

the acquisition of contextual information. The gath-

ered contextual knowledge is then combined with up-

sampled features in the decoders to recover any lost

details. Finally, the binary decoder categorizes pixels

as either lane or non-lane, and the other decoder dis-

tinguishes between individual lane instances. The au-

thors experimented their lane recognition system on

TuSimple (Chang et al., 2019) and Caltech lanes (Aly,

2008) datasets.

Li et al. (2021) developed the Lane-DeepLab

model, enhancing high-deﬁnition map creation for

autonomous driving. Their model architecture incor-

porates a novel attention module added to the Atrous

Spatial Pyramid Pooling (ASPP) module in the en-

coder, enhancing feature extraction, and a Semantic

Embedding Branch (SEB) to merge high and low-

level semantic information for richer feature acquisi-

tion. By leveraging attention mechanisms combined

with contextual semantics, their system adeptly fuses

relevant information to ascertain lane lines with en-

hanced precision. This comprehensive approach en-

ables the model to adapt to and accurately interpret

diverse road situations in complex and dynamically

changing environments. The authors also used the

TuSimple dataset (Chang et al., 2019) and the CU-

Lane (Pan et al., 2018) dataset to demonstrate the ef-

fectiveness of their Lane-DeepLab model.

Lee et al. (2022) proposed a robust lane detection

method using a novel self-attention module called Ex-

panded Self Attention (ESA), optimized for lane de-

tection, that enhances segmentation-based lane detec-

tion by extracting global contextual information. This

ESA module is split into Horizontal (HESA) and Ver-

tical (VESA) components, predicting occluded lane

locations by evaluating lane conﬁdence in both direc-

tions. Their approach, focused on addressing occlu-

sion and challenging lighting conditions, was tested

on three popular datasets: TuSimple (Chang et al.,

2019), CULane (Pan et al., 2018), and BDD100K (Yu

et al., 2020).

Yao et al. (2022) proposed an efﬁcient lane de-

tection approach using a lightweight attention-based

deep neural network, tailored for low memory scenar-

ios. Their architecture comprises two branches: the

Global Context Embedding (GCE) branch for captur-

ing overall lane information, and the Explicit Bound-

ary Regression (EBR) branch, incorporating a Spa-

tial Attention Mechanism (SAM) for precise bound-

ary delineation. The network also employs a Channel

Attention Mechanism (CAM) to prioritize channels

containing target objects. Remarkably, their model

achieved a high performance of 259 frames per sec-

ond (FPS) on an NVIDIA GTX 2070 GPU, with an

input image resolution of 640 x 360. This efﬁciency

was demonstrated on the CULane dataset (Pan et al.,

2018), with the model requiring only 1.57M parame-

ters. The approach was rigorously evaluated on both

the TuSimple (Chang et al., 2019) and CULane (Pan

et al., 2018) datasets

Although these attention mechanisms enhance the

performance of standard models with a minimal in-

crease in parameters, contributing to a slight addi-

tional memory footprint, existing research does not

extensively address their impact on compact CNN

models under stringent memory and inference speed

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

constraints typical of low-spec edge platforms. This

gap is particularly crucial when aiming to maximize

the efﬁciency of standard compact CNN models with

the addition of only a negligible amount of parame-

ters.

3 MODELS FOR LANE

RECOGNITION

3.1 MobileNets

MobileNets, a class of efﬁcient models primarily for

mobile and embedded vision applications, have revo-

lutionized image classiﬁcation by offering a balance

between computational efﬁciency and model accu-

racy (Hanhirova et al., 2018). The architecture lever-

ages depth-wise separable convolutions, signiﬁcantly

reducing the computational burden while maintaining

a competitive edge in performance metrics. Transi-

tioning to MobileNetV2 (Sandler et al., 2018), the in-

troduction of inverted residuals and linear bottlenecks

further optimized the network, enhancing the ﬂow

of information and gradients during training. This

version demonstrates the potential of lightweight yet

powerful models capable of operating in resource-

constrained environments.

Moving to MobileNetV3 (Howard et al., 2019),

the architecture was ﬁne-tuned through automated

Neural architecture search algorithms (Elsken et al.,

2019), embodying a more hardware-aware approach.

This version, known for its improved speed and

efﬁciency, establishes itself as a strong option for

real-time image classiﬁcation tasks. It encourages

progress in incorporating AI into mobile devices

and embedded systems. A notable attribute of Mo-

bileNets is its architectural ﬂexibility, allowing for

seamless integration of individual algorithms such

as attention mechanisms, which can be plugged into

the base architecture, thereby enhancing its func-

tionality and adaptability for diverse computational

paradigms (Sanchez-Iborra and Skarmeta, 2020).

3.2 Channel and Spatial Attention

Mechanisms

The main objective of attention in vision is to emu-

late the human vision cognition process, concentrat-

ing on the crucial patterns present in the input image.

In this work, we decided to evaluate the impact of

attention mechanisms in case of multi-label classiﬁ-

cation applications by selecting two widely-used soft

visual attention techniques: Squeeze-and-Excitation

(SE) (Hu et al., 2018) network and Coordinate Atten-

tion (CA) (Hou et al., 2021). While there are sev-

eral other visual attention techniques available, such

as Spatial Group-wise Enhanced Network(SGE)-

Net (Li et al., 2019), Shufﬂe-Attention Network (SA-

Net) (Zhang and Yang, 2021), and Efﬁcient Channel

Attention (ECA-Net) (Wang et al., 2020), we specif-

ically selected SE and CA as they serve as a base-

line and state-of-the-art, showcasing unique charac-

teristics and have demonstrated effectiveness in con-

volutional neural networks (Guo et al., 2022).

3.2.1 Squeeze-and-Excitation (Channel-Based

Attention)

In the SE (Hu et al., 2018) network, the “squeeze”

phase computes global descriptors for each channel

by aggregating spatial information, which are then

used in the “excitation” phase to learn channel-wise

dependencies and recalibrate channel-wise features,

enhancing the representational capacity of the net-

work (Hu et al., 2018; Guo et al., 2022). The differ-

entiating component in the Squeeze-and-Excitation

equation is the global average pooling which aggre-

gates spatial information across channels.

= σ

H ×W

∑

i=1

∑

j=1

i j

!!!

(1)

Where:

• σ represents the sigmoid activation function

• δ is the ReLU activation function

• W

and W

are the weights of two fully connected

layers.

• X

i j

denotes the input feature map

• H ×W are the spatial dimensions of the input fea-

ture map

3.2.2 Coordinate Attention

The Squeeze-and-Excitation (SE) (Hu et al., 2018)

block initially captures global spatial information via

global pooling and then models the relationships

across channels. However, it overlooks the crucial

aspect of positional information. Coordinate Atten-

tion (CA) (Hou et al., 2021) addresses this limitation

by incorporating positional information into channel

attention, allowing the network to efﬁciently focus on

signiﬁcant large areas with minimal computational re-

sources.

The process within the Coordinate Attention

mechanism involves two distinct phases: coordinate

information embedding and coordinate attention gen-

eration. Initially, two distinct sizes of pooling ker-

nels are employed to process each channel, encoding

Using Attention Mechanisms in Compact CNN Models for Improved Micromobility Safety Through Lane Recognition

Figure 1: A schematic representation of the workﬂow for developing and deploying a neural network model, illustrating the

stages of data collection, model training, optimization, and ﬁnal deployment on a target platform.

along both the horizontal and vertical axes. Following

this, the outputs from these pooling layers are con-

catenated and processed through a shared 1 X 1 con-

volutional transformation function. Subsequently, the

CA mechanism divides the resultant tensor into two

separate tensors. These tensors are then transformed

into attention vectors that align with the horizontal

and vertical dimensions of the input X, each main-

taining the same number of channels.

Unlike traditional channel attention that primarily

focuses on recalibrating channel signiﬁcance, the CA

block extends its functionality by integrating spatial

information encoding. By applying attention simulta-

neously across both the horizontal and vertical planes,

the CA block is adept at spotting the precise locations

of target objects. The differentiating component in the

Coordinate Attention equation is the dual-axis atten-

tion mechanism – horizontal g

(i) and vertical g

( j)

attention weights – that encodes spatial information

along both axes (Eq. 2).

(i, j) = x

(i, j) × g

(i) × g

( j). (2)

where:

• y

(i, j) is the output of the Coordinate Attention

block for the c

channel at position (i, j).

• x

(i, j) represents the input feature map for the c

channel at position (i, j).

• g

(i) and g

( j) are the attention weights for the

horizontal and vertical directions, respectively, at

position (i, j).

4 MICROMOBILITY LANE

RECOGNITION DATASET

(MLRD)

The Micromobility Lane Recognition Dataset

(MLRD) is a novel multi-label classiﬁcation dataset

speciﬁcally designed for lane recognition in mi-

cromobility applications considering the challenges

discussed in the previous section. This section

describes the aspects of our proposed dataset and

explain our approach towards the lane recognition

problem as a multi-label classiﬁcation problem.

The dataset comprises colour images with a pri-

mary focus on three distinct classes: road, sidewalk,

and bike lane, as depicted in Figure 2. These classes

have been carefully chosen to cater to the unique

requirements of micromobility vehicles, such as e-

scooters, for efﬁcient and safe navigation in urban en-

vironments.

A key motivation behind the creation of this

custom dataset is the insufﬁciency of existing

open-source autonomous vehicle datasets, such as

KITTI (Geiger et al., 2013), Cityscapes (Cordts

et al., 2016), DET (Cheng et al., 2019), TuSim-

ple (Chang et al., 2019), LLAMAS (Behrendt and

Soussan, 2019), CurveLanes (Xu et al., 2020), and

nuScenes (Caesar et al., 2020) in addressing the spe-

ciﬁc needs of micromobility applications. Although

these datasets have been instrumental in advancing

computer vision and autonomous vehicle research,

they lack samples of sidewalk and bike lane sections.

The images in such datasets, primarily captured as a

ﬁrst-person view from the car, hinder their applica-

bility in the context of micromobility with the differ-

ent perspective and variable position of e-bikes or e-

scooters.

The images in the MLRD dataset are captured

using a proprietary camera module installed on e-

scooters showcasing streets and their surroundings.

These images are captured from multiple major cities

across Europe and the United States at a resolution of

640x480. The overall dataset consists of 30,244 im-

ages. By combining images from both online sources

and the proprietary camera module, our dataset aims

to provide a comprehensive and diverse collection of

street scenes, with an emphasis on the road, sidewalk,

and bike lane areas. This focus on the most relevant

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

(a) Road Segment (b) Bike lane Segment (c) Sidewalk Segment

Figure 2: Example images from MLRD captured with the e-scooter camera.

areas for micromobility applications will enable the

development of more accurate and efﬁcient models

for lane detection and classiﬁcation, paving the way

for safer and smarter urban mobility solutions. The

MLRD dataset has been made publicly available.

Due to the variety of road and sidewalk structures

and the position (adjacent or separated) of such seg-

ments, annotating frames precisely was a complex

task. Hence, priority when labelling was given to the

road and bike lane classes. The images clearly show-

ing the road or bike lane annotated as 1 for that class

and 0 for the other. Where the bike lane is a part of

the road segment, the image is annotated as both road

and bike lane. Conversely, in cases where the image

clearly consist of a sidewalk or where there is no clar-

ity of a road or bike lane area, it is multi-label anno-

tated as 0 for both the road and bike lane labels. The

MLRD comprises a total of 16,759 samples classiﬁed

as “road”, 5,218 samples as “bike lane”, and 12,510

samples for the indirect class “sidewalk”, where the

labels for both the “road” and “bike lane” classes are

set to zero. The dataset is slightly imbalanced; how-

ever, this issue was addressed by employing an appro-

priate loss function, as explained in the Section 5

The proposed dataset not only focuses on the pri-

mary classes of road, sidewalk, and bike lane but also

incorporates additional labels indicating road mate-

rial such as “asphalt”, “concrete”, and “cobblestone”,

along with labels indicating the time of the day in

the image as “day” and “night”, and weather condi-

tions represented by labels like “sunny”, “cloudy”,

and “rainy”. For the work presented in this paper,

only the the “road” and the “bike lane” classes have

been utilized.

5 EXPERIMENTAL

METHODOLOGY

This section explains the methodology, model de-

ployment and the implementation details of the com-

parative analysis we performed. The experiments

evaluate the impact of channel and spatial atten-

tion with the compact MobileNets (with the width

multiplier α = 0.1) for classiﬁcation by comparing

the performance of ﬁve different variants-Standard

MobileNetV2 (Sandler et al., 2018), MobileNetV2

with channel attention, MobileNetV2 with channel

and spatial attention, standard MobileNetV3 (Howard

et al., 2019) with channel attention and MobileNetV3

with channel and spatial attention on MLRD.

5.1 Experimental Setup and

Hyper-Parameter Details

We used the TensorFlow framework to perform all our

experiments and Weights and Biases MLOps tool to

keep track of all the metrics during the training. For

the training, we used an NVIDIA GEFORCE RTX

4090 GPU with the input image resolution of 224 x

224 and batch size of 32. The initial learning rate

(LR) for Adam optimizer was set to 0.001. The “Re-

duceLROnPlateau” learning rate scheduler was con-

ﬁgured to monitor the validation loss and the “Mod-

elCheckpoint” was adopted to save the best model

with the least validation loss. It had a minimum learn-

ing rate set to a threshold of 1e-6, a reduction factor

of 0.1 and a patience parameter of 10 epochs. We

trained all the models used for the experiments from

scratch on the MLRD without using any pre-trained

weights. We used minimal image augmentations dur-

ing training such as horizontal ﬂips and brightness ad-

justments within the range of 0.2-0.5, while avoid-

ing the augmentation like vertical ﬂips or rotations

to maintain the integrity of the ﬁrst-person micromo-

bility rider perspective in the images. Considering

the class imbalance in MLRD and the objective of

multi-label classiﬁcation, Binary Focal Crossentropy

(BFCE) loss function (Lin et al., 2017) with the com-

monly used weight balancing factor (α) as 0.25 and

the focusing parameter used to compute the focal fac-

tor (γ) as 2.0 were used. All models were trained for

80 epochs.

Using Attention Mechanisms in Compact CNN Models for Improved Micromobility Safety Through Lane Recognition

Figure 3: An illustration of the integration of channel and channel plus spatial attention blocks in the inverted residual block

(a) present in MobileNetV2 (Sandler et al., 2018) and MobileNetV3 (Howard et al., 2019) architectures, as utilized in our

experiments.

5.2 Network Architecture

In our study, we empirically determined that approxi-

mately 100K total trainable parameters, coupled with

an input image resolution of 224x224, allowed for the

successful deployment of our model on the target plat-

form. The detailed description of technical speciﬁca-

tion of the target deployment platform can be found

in section 5.3.

The MobileNetV2 (Sandler et al., 2018) and Mo-

bileNetV3 (Howard et al., 2019) based model archi-

tectures, including the baseline utilized in this re-

search, are derived from the ofﬁcial Keras imple-

mentation. Adhering to the ofﬁcial implementation

guidelines for Squeeze-and-Excitation (SE) (Hu et al.,

2018) and Coordinate Attention (CA) (Hou et al.,

2021), we strategically positioned the SE blocks im-

mediately following the depthwise convolution lay-

ers within the bottleneck modules of the MobileNet

architectures as shown in ﬁgure 3. This placement

enables the SE blocks to recalibrate the features ex-

tracted by depthwise separable convolutions, prior to

their projection through pointwise convolutions (1x1

convolutions) into a higher-dimensional space.

For all model variants in this study, the width mul-

tiplier hyperparameter, which regulates the number of

channels in the bottleneck layers, was ﬁxed at 0.1.

This adjustment did not alter the network’s depth. The

channel reduction ratio for the squeeze operation in

the SE block was set to 16, and for the channel atten-

tion phase in CA, it was established at 32.

To comply with the memory constraints of our tar-

get hardware platform, particularly for MobileNetV3,

we modiﬁed the channels in the bottleneck blocks,

maintaining the original ratio and preserving the in-

tegrity of the original model architecture. Speciﬁ-

cally, we halved the number of output channels and

quartered the expansion factor in each bottleneck

block compared to their original values. The expan-

sion factor is used to expand the number of channels

in the input feature map before applying a depthwise

separable convolution. This expansion allows the net-

work to capture more complex features in a higher-

dimensional space, while maintaining overall compu-

tational efﬁciency.

Furthermore, the neuron count in the ﬁnal Fully

Connected layer was decreased from 1280 to 320,

aligning the model with the computational limitations

of the deployment environment.

5.3 Deployment on the Target Platform

The designated platform for our project is a

STM32H743VI low-spec microcontroller unit, mea-

suring just 1.40in x 1.75in. This unit is equipped

with a 32-bit Arm Cortex-M7 processor that oper-

ates at 480 MHz, supported by 1 MB of Static RAM

and 2 MB of FLASH memory. Despite its diminu-

tive size, it maintains a low power consumption be-

low 150mA, rendering this platform perfectly adapt-

able to the restricted environment of e-mobilities. As

shown in Figure 1, the ﬁnal step in the pipeline in-

volves converting these models to TFLite format to

deploy on the target platform. However, all the ﬂoat32

models mentioned earlier exceeded the required size,

preventing successful deployment. To rectify this,

we compressed the model parameters by transitioning

them from ﬂoat32 to a more microcontroller-friendly

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

Figure 4: A comparative Grad-CAM visualizations illustrating distinct behavioral patterns of different MobileNet model

variants on the MLRD dataset.

int8 precision, utilizing integer quantization using the

Post-Training Quantization technique (Zhang et al.,

2023).

Due to these limitations only the MobileNetV2

model is deployable, as the platform’s ﬁrmware does

not at present accommodate certain tensor operations

present in the MobileNetV3 architecture. Hence,

we formulated our conclusions by conducting exper-

iments in a simulated environment. Slightly differ-

ent behavior is anticipated on the actual platform due

to the quantization process required to compress the

model for deployment.

6 RESULTS AND DISCUSSIONS

Table 1 shows the performance metrics of stan-

dard large MobileNetV2 (MNV2) and MobileNetV3

(MNV3) base models on the MLRD dataset. Due

to the considerably higher number of parameters in

these models compared to their compact counterparts

in Table 2, there remains a signiﬁcant gap in model

performance. The primary purpose of adding atten-

tion mechanisms to the compact models is to mini-

mize this gap as much as possible, aiming to enhance

model efﬁciency without substantially increasing the

computational overhead.

In the comparative analysis of MobileNetV2 and

MobileNetV3 architectures, augmented respectively

with Squeeze-and-Excitation (SE) and Coordinate At-

tention (CA) mechanisms, the experimental outcomes

highlight the utility of channel and spatial information

in micromobility safety application of lane recogni-

tion. These results were particularly noteworthy con-

sidering the challenges of working with fairly low-

resolution images and in a resource-constrained envi-

ronment where learning maximum features with very

minimal computational overhead is essential. The

MNV2 model, integrated with the Coordinate Atten-

tion, demonstrated superior precision scores in classi-

fying both the road and the bike lane classes, with a

signiﬁcant increase in precision ( 8%) and a slight in-

crease (1%) in the average F1 score compared to the

baseline. This precision metric is instrumental in min-

imizing false positives, a crucial aspect for the robust-

ness and reliability of micromobility safety systems.

However, this conﬁguration also manifested in the

highest parameter count, approximately 29K addi-

tional parameters, equivalent to about 100KB of in-

creased memory requirement, marking a 29.55% in-

crease from the baseline. While this conﬁguration

offers potentially the most optimal performance, its

feasibility in low-resource settings is constrained by

the slightly increased computational overhead. The

observed decline in recall metric especially for the

“road” class in this model variant suggests potential

overﬁtting. This overﬁtting implies a slight over-

specialization of the model in recognizing speciﬁc

features, possibly predominant in the bike lane class,

to the detriment of its ability to generalize effectively.

This highlights the critical balance needed between

model complexity and its capacity to perform accu-

rately across diverse real-world scenarios.

In contrast, the MNV2 model integrated with the

Squeeze-and-Excitation attention (channel attention),

with an addition of approximately 11K parameters,

Using Attention Mechanisms in Compact CNN Models for Improved Micromobility Safety Through Lane Recognition

Table 1: Classiﬁcation results between standard (Large) of MobileNets on MLRD.

Models

(Large)

Precision Recall F1 Score

Weighted

Avg F1

Param.

Road Bike lane Road Bike lane Road Bike lane

MobileNetV2 0.96 0.97 0.90 0.86 0.93 0.92 0.93 2.26M

MobileNetV3Large 0.95 0.94 0.92 0.87 0.94 0.91 0.93 2.99M

Table 2: Classiﬁcation results between Attention-Based and Non-Attention-Based compact versions (compact) of MobileNets

on MLRD.

Models

(compact)

Precision Recall F1 Score

Weighted

Avg F1

Param.

Road Bike lane Road Bike lane Road Bike lane

MobileNetV2 (Baseline) 0.86 0.86 0.89 0.80 0.88 0.83 0.86 95.87K

MobileNetV2 + SE 0.87 ↑1% 0.88 ↑2% 0.89 0.79 ↓1% 0.88 0.83 0.87 ↑1% 106.79K

MobileNetV2 + CA 0.94 ↑8% 0.93 ↑7% 0.83 ↓6% 0.79 ↓1% 0.88 0.85 ↑2% 0.87 ↑1% 124.21K

MobileNetV3 (SE) 0.83 ↓3% 0.77 ↓9% 0.74 ↓15% 0.84 ↑4% 0.78 ↓10% 0.80 ↓3% 0.79 ↓7% 109.82K

MobileNetV3 + CA 0.83 ↓3% 0.86 0.85 ↓4% 0.81 ↑1% 0.84 ↓4% 0.83 0.84 ↓2% 88.74K

resulted an slight improvement in precision metric for

both classes and a marginal increase in the overall

F1 score. This enhancement indicates the efﬁcacy of

channel-focused attention mechanisms, like Squeeze-

and-Excitation, in boosting model performance, al-

beit with limitations in capturing the full spectrum of

spatial complexities required for tasks like lane recog-

nition like MobileNetV2 with Coordinate Attention.

For MNV3 models, the variant featuring Coordi-

nate Attention showed consistent performance met-

rics but with a reduced parameter footprint com-

pared to MNV2 models, aligning it more closely with

the resource limitations of low-spec microcontrollers.

However, it is noteworthy that the standard MNV3

model with SE attention was outperformed by its Co-

ordinate Attention-enhanced counterpart, likely due

to SE attention’s limited capability in addressing the

spatial intricacies essential for accurate lane detec-

tion. The integration of Coordinate Attention into

the MNV2 architecture notably enhances its perfor-

mance against both the baseline and SE-boosted vari-

ants. This improvement is attributed to the dual em-

phasis on channel and spatial attention, facilitating a

more comprehensive extraction and analysis of im-

age features. Such an integrated approach is vital for

accurately distinguishing between similar classes like

road and bike lanes, where spatial positioning is often

a key differentiator. Furthermore, the limited impact

of SE attention in these experiments can be attributed

to its primary focus on channel-wise feature recalibra-

tion, lacking the spatial resolution necessary for tasks

that demand an understanding of positional context.

Extending beyond micromobility safety, the im-

plications of these models are signiﬁcant for au-

tonomous vehicle technology, where efﬁcient and

rapid lane recognition is crucial. Implementing these

compact yet effective models in low-spec microcon-

trollers presents a promising path towards developing

cost-effective, high-performance autonomous naviga-

tion systems, offering an efﬁcient alternative to more

computationally intensive segmentation models. This

approach paves the way for the deployment of ac-

curate, yet economically viable autonomous systems,

particularly suited for environments demanding fast

inference and minimal computational load.

Figure 4 presents a Grad-CAM (Selvaraju et al.,

2017) analysis of model variants applied to sam-

ple images from the MLRD dataset. This visualisa-

tion highlights the distinct behavioral patterns of each

model. However, no consistent pattern can be ob-

served in terms of the positioning of objects within

the images to predict the ﬁnal classes.

7 CONCLUSIONS AND FUTURE

WORK

Cities around the world are progressing towards eco-

friendly urban transportation systems, including E-

scooters and E-bikes, to promote sustainable mobil-

ity while reducing trafﬁc, noise, and pollution. Ana-

lyzing the behavior of e-mobility users is critical for

effective governance. However, existing GPS and LI-

DAR systems are limited, necessitating the adoption

of computer vision-based solutions. To bridge this

gap, we contribute a novel lane recognition multi-

label image classiﬁcation dataset speciﬁcally for mi-

cromobilty applications.

Our experiments have demonstrated that while in-

tegrating attention mechanisms into compact CNN

models, such as MobileNetV2 and MobileNetV3, can

yield improvements in precision and overall perfor-

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems

mance, it is not without its challenges. Speciﬁ-

cally, the MobileNetV3 model with channel and spa-

tial attention showed impressive performance, closely

matching the baseline model with signiﬁcantly fewer

parameters. On the other hand, the MobileNetV2

model with channel and spatial attention, while show-

ing improvements in precision, had a decline in re-

call for the “road” class. Similarly, the MobileNetV3

model with channel attention demonstrated an overall

performance deterioration compared to the baseline,

suggesting potential overﬁtting. These observations

indicate that while attention mechanisms can enhance

model accuracy, their integration in compact models

must be approached cautiously, balancing the beneﬁts

against the risks of overﬁtting and increased model

complexity.

In conclusion, we believe that our dataset

(MLRD) serves as a valuable tool for evaluating the

efﬁcacy of MobileNet models with channel and spa-

tial attention mechanisms in enhancing lane recogni-

tion accuracy. These mechanisms hold promise for

deployment in compact models used in micromobil-

ity, especially the MobileNetV2 variants demonstrat-

ing improved F1 scores with a minimal increase in

parameter count. However, the complexities associ-

ated with compression and deployment in micromo-

bility environments can sometimes diminish or off-

set their potential improvements. In the future, we

intend to conduct a more exhaustive comparison in-

volving a wider range of model architectures. We

plan to deploy these models on low-spec target plat-

forms, to evaluate real-world behavior. Additionally,

we intend to explore other more effective and hard-

ware friendly model optimization techniques, such

as structured weight pruning and Quantization Aware

Training.

ACKNOWLEDGEMENTS

This research was conducted with the ﬁnancial sup-

port of Science Foundation Ireland (12/RC/2289

P2)

at Insight the SFI Research Centre for Data Analyt-

ics at Dublin City University and from Luna Sys-

tems. We would like to express our gratitude to Luna

Systems for their invaluable support throughout the

course of this research. Their generosity in providing

us with the essential dataset and the microcontroller

platform was instrumental in the successful comple-

tion of our experiments.

REFERENCES

Aly, M. (2008). Real time detection of lane markers in ur-

ban streets. In 2008 IEEE intelligent vehicles sympo-

sium, pages 7–12. IEEE.

Behrendt, K. and Soussan, R. (2019). Unsupervised la-

beled lane markers using maps. In Proceedings of the

IEEE/CVF international conference on computer vi-

sion workshops, pages 0–0.

Caesar, H., Bankiti, V., Lang, A. H., Vora, S., Liong, V. E.,

Xu, Q., Krishnan, A., Pan, Y., Baldan, G., and Bei-

jbom, O. (2020). nuscenes: A multimodal dataset for

autonomous driving. In Proceedings of the IEEE/CVF

conference on computer vision and pattern recogni-

tion, pages 11621–11631.

Chang, D., Chirakkal, V., Goswami, S., Hasan, M., Jung,

T., Kang, J., Kee, S.-C., Lee, D., and Singh, A. P.

(2019). Multi-lane detection using instance segmen-

tation and attentive voting. In 2019 19th International

conference on control, automation and systems (IC-

CAS), pages 1538–1542. IEEE.

Cheng, W., Luo, H., Yang, W., Yu, L., Chen, S., and Li, W.

(2019). Det: A high-resolution dvs dataset for lane ex-

traction. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition Work-

shops, pages 0–0.

Cicchino, J. B., Kulie, P. E., and McCarthy, M. L. (2021).

Severity of e-scooter rider injuries associated with trip

characteristics. Journal of safety research, 76:256–

261.

Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler,

M., Benenson, R., Franke, U., Roth, S., and Schiele,

B. (2016). The cityscapes dataset for semantic urban

scene understanding. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 3213–3223.

Elsken, T., Metzen, J. H., and Hutter, F. (2019). Neural

architecture search: A survey. The Journal of Machine

Learning Research, 20(1):1997–2017.

Fox, A., Kumar, B. V., Chen, J., and Bai, F. (2017). Multi-

lane pothole detection from crowdsourced undersam-

pled vehicle sensor data. IEEE Transactions on Mo-

bile Computing, 16(12):3417–3430.

Fu, J., Liu, J., Jiang, J., Li, Y., Bao, Y., and Lu, H. (2020).

Scene segmentation with dual relation-aware attention

network. IEEE Transactions on Neural Networks and

Learning Systems, 32(6):2547–2560.

Geiger, A., Lenz, P., Stiller, C., and Urtasun, R. (2013).

Vision meets robotics: The kitti dataset. The Inter-

national Journal of Robotics Research, 32(11):1231–

1237.

GM, V., Pereira, B., and Little, S. (2021). Urban footpath

image dataset to assess pedestrian mobility. In Pro-

ceedings of the 1st International Workshop on Multi-

media Computing for Urban Data, pages 23–30.

Guo, M.-H., Xu, T.-X., Liu, J.-J., Liu, Z.-N., Jiang, P.-T.,

Mu, T.-J., Zhang, S.-H., Martin, R. R., Cheng, M.-M.,

and Hu, S.-M. (2022). Attention mechanisms in com-

puter vision: A survey. Computational visual media,

8(3):331–368.

Using Attention Mechanisms in Compact CNN Models for Improved Micromobility Safety Through Lane Recognition

Hanhirova, J., K

ainen, T., Sepp

a, S., Siekkinen, M.,

Hirvisalo, V., and Yl

a-J

aski, A. (2018). Latency

and throughput characterization of convolutional neu-

ral networks for mobile computer vision. In Proceed-

ings of the 9th ACM Multimedia Systems Conference,

pages 204–215.

Hou, Q., Zhou, D., and Feng, J. (2021). Coordinate atten-

tion for efﬁcient mobile network design. In Proceed-

ings of the IEEE/CVF conference on computer vision

and pattern recognition, pages 13713–13722.

Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B.,

Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V.,

et al. (2019). Searching for mobilenetv3. In Pro-

ceedings of the IEEE/CVF international conference

on computer vision, pages 1314–1324.

Hu, J., Shen, L., and Sun, G. (2018). Squeeze-and-

excitation networks. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 7132–7141.

Lee, M., Lee, J., Lee, D., Kim, W., Hwang, S., and Lee, S.

(2022). Robust lane detection via expanded self atten-

tion. In Proceedings of the IEEE/CVF winter confer-

ence on applications of computer vision, pages 533–

542.

Li, J., Jiang, F., Yang, J., Kong, B., Gogate, M., Dashtipour,

K., and Hussain, A. (2021). Lane-deeplab: Lane se-

mantic segmentation in automatic driving scenarios

for high-deﬁnition maps. Neurocomputing, 465:15–

25.

Li, X., Hu, X., and Yang, J. (2019). Spatial group-wise en-

hance: Improving semantic feature learning in convo-

lutional networks. arXiv preprint arXiv:1905.09646.

Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Doll

ar, P.

(2017). Focal loss for dense object detection. In

Proceedings of the IEEE international conference on

computer vision, pages 2980–2988.

Pan, X., Shi, J., Luo, P., Wang, X., and Tang, X. (2018).

Spatial as deep: Spatial cnn for trafﬁc scene under-

standing. In Proceedings of the AAAI Conference on

Artiﬁcial Intelligence, volume 32.

Sanchez-Iborra, R. and Skarmeta, A. F. (2020). Tinyml-

enabled frugal smart objects: Challenges and opportu-

nities. IEEE Circuits and Systems Magazine, 20(3):4–

18.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2018). Mobilenetv2: Inverted residu-

als and linear bottlenecks. In Proceedings of the IEEE

conference on computer vision and pattern recogni-

tion, pages 4510–4520.

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,

Parikh, D., and Batra, D. (2017). Grad-cam: Visual

explanations from deep networks via gradient-based

localization. In Proceedings of the IEEE international

conference on computer vision, pages 618–626.

Wang, Q., Wu, B., Zhu, P., Li, P., Zuo, W., and Hu, Q.

(2020). Eca-net: Efﬁcient channel attention for deep

convolutional neural networks. In Proceedings of the

IEEE/CVF conference on computer vision and pattern

recognition, pages 11534–11542.

Woo, S., Park, J., Lee, J.-Y., and Kweon, I. S. (2018). Cbam:

Convolutional block attention module. In Proceed-

ings of the European conference on computer vision

(ECCV), pages 3–19.

Xing, Y., Lv, C., Chen, L., Wang, H., Wang, H., Cao,

D., Velenis, E., and Wang, F.-Y. (2018). Advances

in vision-based lane detection: Algorithms, integra-

tion, assessment, and perspectives on acp-based par-

allel vision. IEEE/CAA Journal of Automatica Sinica,

5(3):645–661.

Xu, H., Wang, S., Cai, X., Zhang, W., Liang, X., and Li,

Z. (2020). Curvelane-nas: Unifying lane-sensitive

architecture search and adaptive point blending. In

Computer Vision–ECCV 2020: 16th European Con-

ference, Glasgow, UK, August 23–28, 2020, Proceed-

ings, Part XV 16, pages 689–704. Springer.

Yao, Z., Chen, X., et al. (2022). Efﬁcient lane detection

technique based on lightweight attention deep neural

network. Journal of Advanced Transportation, 2022.

Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F.,

Madhavan, V., and Darrell, T. (2020). Bdd100k: A

diverse driving dataset for heterogeneous multitask

learning. In Proceedings of the IEEE/CVF conference

on computer vision and pattern recognition, pages

2636–2645.

Zakaria, N. J., Shapiai, M. I., Ghani, R. A., Yasin, M.,

Ibrahim, M. Z., and Wahid, N. (2023). Lane detection

in autonomous vehicles: A systematic review. IEEE

Access.

Zhang, J., Zhou, Y., and Saab, R. (2023). Post-training

quantization for neural networks with provable guar-

antees. SIAM Journal on Mathematics of Data Sci-

ence, 5(2):373–399.

Zhang, L., Jiang, F., Kong, B., Yang, J., and Wang, C.

(2021). Real-time lane detection by using biologically

inspired attention mechanism to learn contextual in-

formation. Cognitive Computation, 13:1333–1344.

Zhang, Q.-L. and Yang, Y.-B. (2021). Sa-net: Shuf-

ﬂe attention for deep convolutional neural networks.

In ICASSP 2021-2021 IEEE International Confer-

ence on Acoustics, Speech and Signal Processing

(ICASSP), pages 2235–2239. IEEE.

Zhu, X., Cheng, D., Zhang, Z., Lin, S., and Dai, J. (2019).

An empirical study of spatial attention mechanisms in

deep networks. In Proceedings of the IEEE/CVF inter-

national conference on computer vision, pages 6688–

6697.

VEHITS 2024 - 10th International Conference on Vehicle Technology and Intelligent Transport Systems