Edge‑Enabled Sensor Fusion and Hybrid Machine Learning

Framework for Real‑Time Smart Parking Detection and Scalable

Occupancy Prediction

S. V. Dharani Kumar

, Swaraj Satish Kadam

, Girija M. S.

, P. Mathiyalagan

D. B. K. Kamesh

and Tamilselvi E.

Department of Biomedical Engineering, GRT Institute of Engineering & Technology, GRT Mahalakshmi Nagar, Tiruttani,

Tiruvallur dist - 631209, Tamil Nadu, India

Department of Electrical Engineering, Dr. D Y Patil Institute of Technology, Pimpri, Pune 411018, Dr. D. Y. Patil Dnyan

Prasad University, Pune, India

Department of Computer Science and Design, R.M.K. Engineering College, Kavaraipettai, Tamil Nadu, India

Department of Mechanical Engineering, J.J. College of Engineering and Technology, Tiruchirappalli, Tamil Nadu, India

Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad‑500043, Telangana, India

Department of ECE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Keywords: Sensor Fusion, Edge Computing, Smart Parking, Occupancy Prediction, Hybrid Machine Learning.

Abstract: Urban mobility is a real challenge in terms of inefficient parking spaces detection and occupation in densely

populated areas. This paper presents an edge-enabled, sensor-integrated, hybrid machine learning framework

targeted for real-time smart parking detection and predictive occupancy analytics. Rather than the

conventional single-sensor-based and static-learning-dependent models, our system integrates ultrasonic,

infrared, magnetic, and vision sensors to achieve the robust detection under various environmental situations.

CNN, LSTM and XGBoost are equally combined to accurately make temporal and spatial predictions, and

edge processing is energy-efficient to reduce latency and guarantee privacy protection. Learning mechanisms

are adaptive and enabling the system to be trained incrementally based on real-time feedback and environment

cues, thus scalable for city implementation. This unified method not only enhances the detection performance

and energy saving but also is a stepping stone towards intelligent city infrastructure with a balance of

responsiveness and sustainability.

1 INTRODUCTION

The fast urbanization of contemporary cities has

resulted in a high demand for mobility and efficient

infrastructure, with smart parking systems being one

kind of crucial system to tackle traffic congestion

and reduce environmental impact. As urban

populations increase and the volume of traffic rises,

traditional parking methodologies are being left

behind, with most of these methods relying on manual

observation or on simplistic counting based

approaches (that are not accurate, scalable and

versatile). Sensing technologies combined with

machine learning has risen promising solution to cope

with these drawbacks.

Existing works normally rely on one kind of

single sensor or centralized cloud-based model which

are vulnerable to data latency, low reliability in

diverse environment and high area coverage cost. In

addition, most of these systems do not respond to the

dynamic nature of factors including evolving weather

conditions, traffic congestion levels, and user habits

in real time. This causes suboptimal use of space,

disappointment to users and energy waste.

To solve these problems, this paper introduces the

concept of a next-generation smart parking

framework based on multisensor fusion including

infrared, ultrasonic, magnetic and visual sensor data.

These heterogeneous inputs areprocessed in a local

pipeline of edge computing nodes which permits real-

time decision making while preserving user data

privacy. The model is designed with a combination of

Kumar, S. V. D., Kadam, S. S., S., G. M., Mathiyalagan, P., Kamesh, D. B. K. and E., T.

Edgeâ

SEnabled Sensor Fusion and Hybrid Machine Learning Framework for Realâ

STime Smart Parking Detection and Scalable Occupancy Prediction.

DOI: 10.5220/0013857100004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

34-40

ISBN: 978-989-758-777-1

Convolutional Neural Network (CNN), Long Short

Term Memory (LSTM), and gradient boosting

method (XGBoost), to complement detection with

spatial context, and to make temporal prediction, and

to refine decision.

To scale across heterogeneous city zones, we also

design our method to self-adapt online and leverage

Model Updating and Feedback Loops guaranteeing

no-retraining from scratch. Such smart infrastructure

does not only enable precise and real-time parking

space surveillance, but also serves as the basis for

predictive analytics regarding future space

occupation. Through the resolution of the underlying

problems in previous methods, the sparse SIP

framework provides a technical and sustainable

pipeline that is compatible with the future smart and

responsive urban environments.

1.1 Problem Statement

Although many urban mobility solutions have been

already developed, the in real-time detection and

accurate forecasting of parking spaces occupation

represents still an unsolved problem in smart city

environments. The current parking management

systems are commonly built based on single sensor

dependence, centralized processing, and non-

adaptive machine learning methods. However, such

methods exhibit poor performance when the scenario

is dynamic and latency is large for decision-making,

and are difficult to adapt for large and complex urban

infrastructures.

In addition, currently the existing systems do not

have an access to heterogeneous sources, and they do

not perform real-time processing in a scalable

manner. As a result, low usability and user experience

in using the systems, high traffic congestion and

superfluous fuel consumption. Privacy violations can

also occur when sensitive visual or location

information is sent to remote server(s) for processing

and stored without sufficient safeguards. These

issues motivate the urgency to design a scalable, low-

latency, energy-efficient, and privacy-aware

approach that processes heterogeneous sensor data

and evolves with various urban scenes.

This paper fills this gap by proposing an edge-

based, sensor-fused, hybrid machine learning

framework for improvement of real-time parking

detection and prediction accuracy with scalability and

adaptation, and user data privacy with the state-of-

the-art.

2 LITERATURE SURVEY

Intelligent parking systems have experienced a

substantial growth with the emergence of internet of

things (IoT), sensor networks, and machine learning.

Conventional parking systems were generally stand-

alone and manual with little real-time querying or

predictive ability. Recent work has attempted to

overcome these limitations by employing machine

learning and sensor fusion but crucial challenges

remain.

Pore and Nemade (2025) developed a vision-

based empty-parking-slot detection system using

deep learning without sensor fusion, which

underperformed in low light and with obstructions,

however. Kasera and Acharjee \cite{8851050} also

have used LSTM model for occupancy prediction,

effective for short-term but ineffective for the real-

time visual noise and contextual sensor data.

In a systematic review by Navpreet et al. (2025)

has emphasized the necessity of hybrid, or ensemble

models because their approach primarily based on a

single model, such as a CNN or SVM, cannot provide

the necessary flexibility and precision in the real-time

urban condition as needed. Furthermore, Sahoo et al.

(2021) presented smart parking using IoT, although

they admitted high latency on the cloud and weak

security because of both cloud-centric and poor

integration of edge computing.

Enríquez et al. (2024), developed a fog

computing-oriented architecture for mitigating data

processing latency; however, they employed

incomplete sensor input (e.g., only infrared or just

magnetic) to decrease the robustness of detection

when the environment is variable. Mehmeti and

Stojanovski (2025) additionally reviewed sensor

placement techniques in combination with machine

learning, but the work did not have an end-to-end

real-time implementation, specifically in dense urban

environments.

Additionally, Yang et al. (2019) developed a deep

learning based model and applied spatio-temporal

data for prediction, but its performance was only

assessed within simulated setting, and several issues

remained to achieve real-world deployment. Data

privacy and power efficiency challenges still persist

within many of these systems, where most techniques

do not leverage edge computing (to avoid reliance on

high bandwidth, cloud-centric infrastructures to

realize gains) (Mejía-Muñoz et al., 2024).

The trend towards more intelligent and energy-

conscious architectures is evident but cost-effective,

distributed, scalable, and privacy-complying

solutions are still developing. Only a small number of

Edgeâ

SEnabled Sensor Fusion and Hybrid Machine Learning Framework for Realâ

STime Smart Parking Detection and Scalable

Occupancy Prediction

models offer adaptive or self-learning capabilities

that allows improving predictions based on live

feedback from the environment.

This literature work collectively indicates the

following possible research gap: the demand for a

scaleable, adaptive, multi-sensor and edge-capable

system for real-time occupancy detection and

prediction. The introduced framework in this study

fills this gap by leveraging heterogeneous sensing and

hybrid AI models (CNN-LSTM-XGBoost), thus

generating urban parking infrastructure, which is

more robust, accurate, privacy-aware and scalable for

the future.

3 METHODOLOGY

The proposed approach is built upon the fusion of

multimodal sensor data based on the brought together

edge processing and a combination of machine

learning (ML) architecture, enabling accurate

parking space occupancy detection and prediction.

The nodes include multi-sensor devices installed in

different parking areas. Figure 1 explains

deployment of sensor hardware in a smart parking

bay including IR, ultrasonic, magnetic, camera and

their spatial location. That is, all the nodes are

embedded with the integrated devices such as

infrared (IR) sensors, ultrasonic sensors, magnetic

detectors, and cameras. This heterogeneous sensor

layout provides fault tolerance and robustness to

varying environmental conditions, like light

variation, occlusions or weather effects. We locally

pre-process sensor data at the edge node through

trivial filtering for noise suppression and

normalization - yet to standardize the quality of the

input between different modality. Figure 1 illustrates

the smart parking sensor layout.

Figure 1: Smart Parking Sensor Layout.

After data is collected and preprocessed, it is fed

into a \emph{multi-stream processing pipeline}. To

start, each sensor input is first fed through the model

designed for that modality’s data. For instance visual

data flows through a Convolutional Neural Network

(CNN) for spatial feature extraction and evolutionary

sequences from occupancy logs or sensor time-

stamps that are modelled with Long Short-Term

Memory (LSTM) networks to characterize patterns of

occupancy over time. At the same time, the readings

of the magnetic and ultrasonic sensors are input to an

XGBoost classifier to perform fine-grained

Sensor Data

Collection

Preprocessing

at Edge Node

Parallel Data

Streams to ML

Models

•CNN

•LSTM

• XGBoost

Decision-Level

Fusion

Real-Time

Occupancy

Status +

Prediction

Feedback Loop

Model Update &

Continuous

Learning

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

classification with respect to certain predefined

occupancy thresholds. Table 1 compares the

effectiveness of the individual ML models in the

hybrid framework. Outputs of all three models are

combined at the decision level utilising a weighted

ensemble approach, with weights being adapted on-

the-fly by considering sensor reli- ability metrics as

well as the contextual feedback.

Table 1: Machine Learning Models Used.

Model

Input

Task Strength

CNN

Image

data

Spatial

detection

Excellent

visual feature

capture

LSTM

Time-

series

data

Temporal

pattern

forecasting

Captures

historical

trends

XGBo

ost

Tabular

sensor

data

Classification

of occupancy

High

interpretability,

fast

To minimize the latency and be real-time

responsive, the complete model ensemble is

implemented onto edge devices like Raspberry Pi or

Nvidia Jetson Nano. These edge nodes have the

inputs processed, make local predictions and only

pass aggregated insights onto the central server

which helps in saving bandwidth and maintaining

data privacy. For a more scalable system, a cloud-

based orchestrator that runs on container-based

architecture (Docker and Kubernetes) is applied for

the system which enables the system to perform

model updating, collecting feedback for continuous

learning and system monitoring.

Table 2 shows the

sensor specifications and roles.

Table 2: Sensor Specifications and Roles.

Sensor Type Specification Role in Framework Deployment Location

Ultrasonic 2cm–400cm range, 40kHz Detects presence of vehicle Ground-level

Infrared (IR) <2m detection range Motion and heat sensing Wall-mounted

Magnetic 3-axis, ±8 Gauss Detects metallic object changes Embedded in pavement

Camera (RGB) 720p/1080p Visual image capture for CNN processing Overhead mount

Additionally, the system provides feedback

learning capability, in which real-world parking

outcomes (e.g., actual occupancy as confirmed by a

user app or city traffic feed) are utilized to refine the

model over time. This RL-loop ensures that the

framework continuously learns to adapt to changes in

patterns, e.g., seasonal shifts in demand, degradation

of sensor readings, or changes in infrastructure.

Power efficiency is achieved by activating sensors per

occupancy status across the network, and utilizing

sleep-wake scheduling for low-activity environment

making it more energy efficient for longer term

deployments.

The proposed methodology is tested with

synthetic as well as real datasets from a smart city

pilot zones. Performance metrics such as the

detection accuracy, prediction error, latency and

power consumption are employed to demonstrate the

efficiency of the proposed framework. This

distributed and adaptable architecture provides a

solution in the present context of smart parking and

offers an evolutionary and sustainable approach that

will fit the future urban mobility innovation.

4 RESULTS AND DISCUSSION

The proposed approach was further tested with real-

world urban data collected in three test areas over 30-

50 parking spaces with different sensor setups and

environmental conditions. Figure 3: Comparison of

average prediction latency among edge-enabled,

traditional ML, and cloud-only systems.

Table 3

gives the Energy Consumption Comparison.

Table 3: Energy Consumption Comparison.

System

Variant

Avg Power

Usage (W)

Battery

Life (Est.

hrs

)

Remarks

Edge-

Enabled

Framewor

3.8 48

Energy-

efficient,

real-time

Cloud-Based

6.5 28

High

transmission

energ

We have evaluated performance based on

parameters such as prediction precision, occupancy

detection accuracy, model latency, energy

Edgeâ

SEnabled Sensor Fusion and Hybrid Machine Learning Framework for Realâ

STime Smart Parking Detection and Scalable

Occupancy Prediction

consumption and time adaptability. Figure 2

illustrates the comparison of detection accuracy

between single machine learning models and the

hybrid ensemble model in the testing dataset.

Table 4

tabulates the Evaluation Metrics and Results.

Table 4: Evaluation Metrics and Results.

Metric

Propose

d Model

Traditiona

l ML

Cloud-

Only

Approach

Detection

Accuracy

(

)

95.8 86.3 88.1

Prediction

RMSE

0.09 0.17 0.14

Avg

Latency

(ms)

210 620 840

Figure 2: Detection Accuracy Comparison.

By comparison to standalone models, the hybrid

ensemble model with CNN for visual feature

extraction, LSTM for temporal occupancy trend

modeling, and XGBoost for sensor-based

classification achieved the highest accuracy. In

detection, the performance of the CNN component is

89.6% on average over clear visual conditions,

however, it degraded under low illumination and

occluded view. However, taking advantage of

ensemble learning of ultrasonic and magnetic sensor

data, the overall detection accuracy reached 95.8%,

which indicated the robustness of real-world

deployment.

Figure 3 graphs the prediction latency

comparison.

Figure 3: Prediction Latency Comparison.

Performance of prediction was tested with Mean

Absolute Error (MAE) and Root Mean Square Error

(RMSE) metrics. The occupation prediction of the

LSTM had an RMSE and MAE of about 0.09 and

0.06, respectively, on peak hours, which outperforms

traditional ARIMAbased baselines with an RMSE of

0.16. Particularly, the XGBoost model was well-

suited for enhancing prediction accuracy under non-

standard usage patterns (heavily used special events

or rainy days), where temporal models only were less

effective.

Table 5 gives the environmental robustness

test.

Table 5: Environmental Robustness Test.

Condition

Accurac

y (%)

Most

Affected

Senso

Compensatin

g Sensor

Clear

96.7 None N/A

Rainy

Weathe

91.2 Camera Magnetic

Night-

Time

93.5 IR Camera

Dust/Obs

truction

90.6 Ultrasonic IR

Low latency measurements showed that the edge-

deployed inference pipeline operated with an average

response time ranging from 180–250 milliseconds,

making real-time inferences possible without

requiring high bandwidth connections to the cloud.

Figure 4 shows the system accuracy for various

environmental conditions: rain, night, and visual

obstructions. By contrast to the conventional cloud-

based parking systems that exhibited a 600–900 ms

response time for network attributed delays, our

system presented a significant increase in

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

responsiveness required for a user-facing application

or a traffic regulation system.

Figure 4: Environmental Conditions Vs Accuracy.

From an energy point of view, the architecture

minimised power consumption through contextual

sensor activation. As a result, it achieved a 34%

reduction in energy consumption compared with

systems with fully active sensors. The edge devices

functioned under power-constrained situations with

no noticeable performance deterioration, indicating

that they could be deployed in both solar and battery-

powered configurations.

The conversation also uncovered that adding a

feedback loop for model retraining enhanced long-

term generalization. For example, despite being

trained on pre-learned sane traffic, if artificial

anomalies were fed into the learning system, say by

increasing car inflow into the system, the learning

process accounted for it and only after 3 learning

cycles, the pre-posterior learning would absorb real-

time responding and learning. In a 30-day run, models

using feedback learning increased predictive

accuracy by a further 3.4%, thus confirming the value

of adaptive intelligence in dynamic urban

environments.

While the successful outcomes were

encouraging, there were a few issues that were

observed. The visual data was less reliable when the

rainfall was heavy, with CNN improvements. Figure

5 shows the average power consumption comparison

between edge-enabled, cloud-based, and always-on

smart parking systems.

But the advantage of non-visual sensors being

very reliable. Moreover, preliminary deployment

demonstrated that edge nodes needed manual

calibration to tune sensor fusion weights, a process

that might be further automated with meta-learning in

future designs.

Figure 5: Power Consumption Comparison.

Finally, our results validate the importance of

sensor fusion with hybrid machine learning and edge

computing for both the performance and efficiency

of smart parking systems. The model presented not

only decreases the computational cost and latency,

but also guarantees adaptability, privacy and

scalability, which are the prerequisites in deploying

such system in today’s smart cities. Learnings from

this evaluation provide a road map for applying

similar methods to other smart infrastructure

domains, e.g., traffic flow control, smart charging

stations, and vehicle coordination.

5 CONCLUSIONS

This study introduced a large-scale dynamic smart

parking framework which combined edge-enabled

sensor fusion and hybrid-ML architecture in real-time

in order to meet the increasing need of intelligent

urban infrastructure. The combination of visual,

inertia, sonic and magnetic sensor led increased the

system robustness in dynamic environments. An

ensemble model with three basic models

(CNN+LSTM+XGBoost) was proposed to improve

the occupancy detection and prediction compared to

the classical ones.

Edge computing greatly reduced latency, kept

user privacy and contributed to energy efficiency due

to local processing as well as context-aware sensor

activation. Moreover, we integrated a feedback-based

learning mechanism to ensure that the system was

continuously evolving in real-time to match urban

Edgeâ

SEnabled Sensor Fusion and Hybrid Machine Learning Framework for Realâ

STime Smart Parking Detection and Scalable

Occupancy Prediction

dynamics - proving the scalability and relevance for

large-scale deployment in the city.

The results of the experiments confirmed that the

framework is effective in practice and that it has the

potential to enable solutions for less congested

streets, less CO2 emissions caused by unnecessary

parking search and city-led smart mobility services

that are citizen driven. There are still some issues on

which future improvements of the system will focus;

namely, under extreme weather conditions but, given

the modular and adaptive structure of the system,

there is potential to further build in future through

self-calibration and reinforcement learning

mechanisms.

In conclusion, this work paves the way for

intelligent, secure, scalable, and green-sensitive

parking management systems, which are paramount

to the next generation Smart City.

REFERENCES

Enríquez, J. M., Mejía-Muñoz, J. M., Bravo, G., & Cruz-

Mejía, O. (2024). Smart parking: Enhancing urban

mobility with fog computing and machine learning.

Applied System Innovation, 7(3), 52. MDPI

Enríquez, J. M., Mejía-Muñoz, J. M., Bravo, G., & Cruz-

Mejía, O. (2024). Smart parking: Enhancing urban

mobility with fog computing and machine learning.

Applied System Innovation, 7(3), 52. MDPI

Kasera, R. K., & Acharjee, T. (2025). Parking slot

occupancy prediction using LSTM. Innovations in

Systems and Software Engineering, 21, 65–77.

SpringerLink

Kasera, R. K., & Acharjee, T. (2025). Parking slot

occupancy prediction using LSTM. Innovations in

Systems and Software Engineering, 21, 65–77.

SpringerLink

Kasera, R. K., & Acharjee, T. (2025). Parking slot

occupancy prediction using LSTM. Innovations in

Systems and Software Engineering, 21, 65–77.

SpringerLink

Mehmeti, B., & Stojanovski, N. (2025). Machine learning

models and mathematical approaches for improving

IoT-based smart parking systems. Sensors, 25(7), 2065.

MDPI

Mehmeti, B., & Stojanovski, N. (2025). Machine learning

models and mathematical approaches for improving

IoT-based smart parking systems. Sensors, 25(7), 2065.

MDPI

Mejía-Muñoz, J. M., Enríquez, J. M., Bravo, G., & Cruz-

Mejía, O. (2024). Smart parking: Enhancing urban

mobility with fog computing and machine learning.

Applied System Innovation, 7(3), 52. MDPI

Mejía-Muñoz, J. M., Enríquez, J. M., Bravo, G., & Cruz-

Mejía, O. (2024). Smart parking: Enhancing urban

mobility with fog computing and machine learning.

Applied System Innovation, 7(3), 52. MDPI

Navpreet, R., Rani, R., & Roul, R. K. (2025). State-of-the-

art machine learning and deep learning techniques for

parking space classification: A systematic review.

Archives of Computational Methods in Engineering.

SpringerLink

Navpreet, R., Rani, R., & Roul, R. K. (2025). State-of-the-

art machine learning and deep learning techniques for

parking space classification: A systematic review.

Archives of Computational Methods in Engineering.

Navpreet, R., Rani, R., & Roul, R. K. (2025). State-of-the-

art machine learning and deep learning techniques for

parking space classification: A systematic review.

Archives of Computational Methods in Engineering.

SpringerLink

Pore, A. A., & Nemade, P. D. (2025). A novel approach to

detect parking space occupancy for efficient urban

management. Civil Engineering Journal, 11(2).

Pore, A. A., & Nemade, P. D. (2025). A novel approach to

detect parking space occupancy for efficient urban

management. Civil Engineering Journal, 11(2).

Pore, A. A., & Nemade, P. D. (2025). A novel approach to

detect parking space occupancy for efficient urban

management. Civil Engineering Journal, 11(2).

Sahoo, B. B. P., Shahjad, & Tanwar, P. S. (2021). Real-time

smart parking: Challenges and solution using machine

learning and IoT. International Journal of Scientific

Research in Computer Science, Engineering and

Information Technology, 7(2). IJSRCSEIT

Sahoo, B. B. P., Shahjad, & Tanwar, P. S. (2021). Real-time

smart parking: Challenges and solution using machine

learning and IoT. International Journal of Scientific

Research in Computer Science, Engineering and

Information Technology, 7(2). IJSRCSEIT

Yang, S., Ma, W., Pi, X., & Qian, S. (2019). A deep

learning approach to real-time parking occupancy

prediction in transportation networks incorporating

multiple spatio-temporal data sources. Transportation

Research Part C: Emerging Technologies, 107, 248–

265. SpringerLink

Yang, S., Ma, W., Pi, X., & Qian, S. (2019). A deep

learning approach to real-time parking occupancy

prediction in transportation networks incorporating

multiple spatio-temporal data sources. Transportation

Research Part C: Emerging Technologies, 107, 248–

265. SpringerLink

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES