Received 18 January 2025, accepted 18 March 2025, date of publication 25 March 2025, date of current version 4 April 2025.

Digital Object Identifier 10.1109/ACCESS.2025.3553854

AI-Driven Smart Shopping Carts With Real-Time

Tracking and Inventory Forecasting for
Enhanced Retail Efficiency
MUHAMMAD IMRAN ZULFIQAR 1 , AYESHA KHALID 2, ABUBAKR SIDDIG 3,

MUHAMMAD JUNAID NAWAZ4 , AND SALIM SAAY 5

1 Schoolof Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
2 Department of Computer Science and Information Technology, University of Jhang, Jhang 35200, Pakistan
3 Schoolof Computer Science, Technological University Dublin, Dublin 7, D07 EWV4 Ireland
4 Department of Physical and Numerical Sciences, Qurtuba University of Science and Information Technology, Dera Ismail Khan 29111, Pakistan
5 Department of Computer Science and Information Systems, University of Limerick, Limerick, V94 T9PX Ireland

Corresponding author: Abubakr Siddig (abubakr.siddig@tudublin.ie)

ABSTRACT This paper presents an AI-driven smart shopping cart system designed to enhance retail
efficiency and customer experience through real-time data analytics and machine learning. Traditional
shopping carts lack capabilities for adaptive tracking, inventory management, and personalized customer
interaction. Our system addresses these gaps with a multi-layered architecture that integrates person-
specific tracking, reinforcement learning (RL) for navigation, and Long Short-Term Memory (LSTM)
networks for demand forecasting, alongside seamless Point-of-Sale (POS) integration for automated billing.
The architecture comprises real-time data capture, edge computing for low-latency decisions, and cloud
processing for customer profiling and inventory management. Experimental results demonstrate notable
improvements in tracking accuracy, navigation efficiency, inventory forecasting, and customer satisfaction,
highlighting AI’s transformative potential in retail.

INDEX TERMS Artificial intelligence, machine learning, smart shopping cart, retail, inventory
management, personalized shopping, autonomous systems.

I. INTRODUCTION Traditional shopping carts, limited in functionality, do not

The retail industry is undergoing a significant transformation, meet the increasing demands of modern retail environments,
propelled by advanced technologies, including Artificial lacking capabilities for real-time navigation optimization,
Intelligence (AI), Machine Learning (ML), and edge com- personalized product recommendations, and streamlined
puting [1], [2], [3], [4]. These technologies are instrumental checkout [6], [7], [8]. Manual inventory management often
in addressing essential retail challenges, such as enhancing fails to match fluctuating customer demand, resulting
customer experience, improving inventory accuracy, and in frequent stock shortages or excessive stockpiles. The
automating checkout processes. An emerging innovation proposed AI-driven smart shopping cart overcomes these
within this ecosystem is the AI-powered smart shopping cart, limitations by integrating advanced AI algorithms, real-
a system poised to streamline the entire in-store shopping time data processing, and RL techniques, creating a fully
process—from product selection to automated billing— autonomous, adaptive solution to modern retail challenges.
thereby providing retailers and customers with an efficient, This research is based on the following question:
personalized, and adaptive shopping experience [5]. • How can an AI-driven smart shopping cart system
be developed to address the limitations of traditional
carts, enabling real-time tracking, optimized navigation,
The associate editor coordinating the review of this manuscript and personalized product recommendations, and accurate
approving it for publication was Junho Hong . inventory forecasting to enhance retail efficiency?
2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
55576 For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 13, 2025
M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

In this study, we aim to introduce a novel AI-powered smart TABLE 1. Symbols and notations used in this research.
shopping cart system engineered to optimize customer expe-
rience and operational efficiency. This system uses a multi-
layered architecture: real-time data collection is achieved
through hardware components, low-latency decision-making
through edge computing, and intensive computational tasks
such as demand forecasting and customer profiling are
handled via cloud infrastructure. Person-specific tracking is
achieved through Red, Green, Blue, and Depth (RGB-D)
cameras, ultrasonic sensors, and Kalman filters integrated
with re-identification (Re-ID) models, ensuring uninterrupted
and accurate customer tracking. RL algorithms dynamically
adjust the cart’s navigation path based on store layout and
the experimental setup, Section VI presents the results,
real-time obstacle detection, ensuring efficient movement
and Section VII concludes with key findings and future
throughout the store.
directions.
Moreover, LSTM networks are utilized to forecast product
demand and optimize inventory levels, thus minimizing both
stockouts and overstock situations and reducing operational II. BACKGROUND AND RELATED WORK
costs. Our system seamlessly integrates with POS sys- The swift progress in AI and ML has significantly impacted
tems, automating billing and enabling contactless payments the retail industry, introducing solutions that enhance
for improved customer convenience. Collectively, these operational efficiency, improve customer satisfaction, and
innovations contribute to an AI-driven retail ecosystem streamline inventory management [9], [10]. AI is widely
that significantly improves operational efficiency, inventory applied in decision-making systems, while ML models excel
management, and customer satisfaction. in predictive tasks, learning from both historical and real-
The main contributions of this paper are as follows: time data. The integration of these technologies has enabled
• We propose an AI-powered smart shopping cart system
the development of sophisticated systems that automate
that integrates real-time data acquisition, edge comput- various retail processes, such as customer interaction, product
ing, and cloud processing, thereby greatly enhancing recommendation, and navigation [11]. Smart shopping carts
shopping experience and operational efficiency in retail equipped with AI and ML have, in particular, attracted
environments. attention for their potential to transform the traditional
• We develop a precise person-specific tracking algorithm shopping experience.
leveraging RGB-D cameras, Kalman filters, and Re-ID
models based on cosine similarity, ensuring robust, real- A. SMART SHOPPING CARTS
time tracking of customers throughout their in-store Smart shopping carts are engineered to automate and improve
journey. different aspects of in-store shopping, featuring capabilities
• The system implements RL-based dynamic navigation such as product scanning, navigation, and billing automation.
optimization, enabling adaptive pathfinding that adjusts Early implementations, like Amazon Dash Cart [12] and
to store layouts and obstacles in real time, improving Caper’s Smart Cart [13], focused on checkout simplification
navigation efficiency. by eliminating manual product scanning through RFID tech-
• We apply LSTM networks for accurate demand forecast- nology, barcode scanners, and computer vision algorithms.
ing, facilitating optimal inventory management while However, these systems lacked advanced decision-making
minimizing stockouts and overstock, thus significantly capabilities and personalized customer interactions achiev-
reducing associated operational costs. able through sophisticated ML models and AI-powered
• Our system integrates seamlessly with POS systems, systems [14].
automating billing, supporting real-time product scan- Recent studies have explored enhancements to smart
ning, and enabling contactless payments to streamline shopping carts through more advanced AI and ML models.
checkout processes and enhance customer convenience. For example, AI-driven shopping systems have been devel-
• An extensive experimental evaluation is conducted oped for real-time product recommendations, automated
in both simulated and real-world retail environments, pathfinding, and predictive analytics for inventory manage-
demonstrating marked improvements in customer satis- ment [15]. While these innovations mark progress, they are
faction, operational efficiency, and inventory accuracy. often limited to isolated functions such as automated billing
The key notations used throughout this manuscript are or static navigation, without fully exploiting AI’s potential
summarized in Table 1. for adaptive, real-time decision-making in dynamic retail
The paper is organized as follows: Section II reviews environments.
related work, Section III describes the system architecture, Our proposed AI-powered smart shopping cart system
Section IV presents research method, Section V details advances these efforts by integrating RL algorithms for

VOLUME 13, 2025 55577

 M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

optimized in-store navigation and LSTM networks for with real-time customer tracking and interaction. Using
forecasting-based inventory management. This integration person-specific tracking algorithms, the system reliably
enables the cart to autonomously adapt to changing store lay- follows customers throughout the store, while RL dynam-
outs, customer preferences, and product demand, delivering ically optimizes navigation based on store layout and
a personalized and efficient shopping experience. customer behavior. By merging AI-driven personalization
with autonomous navigation, our system delivers an intuitive
B. AI AND ML IN THE RETAIL SECTOR
and seamless shopping experience, elevating customer
satisfaction and operational efficiency.
AI and ML applications in retail have shown substantial
potential in optimizing both customer experiences and
operational processes [16], [17]. AI is applied across D. CHALLENGES IN AI-DRIVEN RETAIL SYSTEMS
various retail functions, such as customer behavior analysis, Despite advancements in AI and ML, challenges remain in
demand forecasting, and recommendation generation [18]. deploying AI-driven retail systems. One primary challenge
ML techniques, especially deep learning and RL, play a is integrating AI models with existing retail infrastructure,
crucial role in enabling these systems to continuously learn such as POS systems, inventory databases, and customer
from extensive datasets and improve performance over time. relationship management platforms [11]. Seamless com-
Convolutional Neural Networks (CNNs), for instance, are munication among these systems is essential to achieve
widely used in recommendation systems to predict customer real-time synchronization of customer data, inventory levels,
preferences based on historical purchasing patterns and real- and product information.
time interactions [14]. Data privacy and security are also critical, as AI-driven
RL, in particular, has proven effective for path opti- systems often collect and process substantial volumes of
mization, notably within autonomous robotic systems in customer data [25]. Retailers must secure this data and ensure
warehouses and retail stores [19], [20]. By learning from compliance with privacy regulations. Additionally, scalability
the environment, RL models can continuously optimize poses a challenge, as AI systems must manage interactions
navigation paths, reducing travel time and avoiding obstacles, with hundreds or thousands of customers concurrently
thereby improving overall operational efficiency. Despite without compromising performance.
these advancements, the application of these techniques to This study addresses these challenges by developing a
smart shopping carts remains limited, with research often smart shopping cart system that integrates with existing retail
focusing on discrete aspects such as inventory management infrastructure while ensuring data privacy through encrypted
or customer interaction, rather than developing an integrated communications and role-based access control. The system is
system. also highly scalable, capable of supporting multiple carts and
This work addresses this gap by presenting a unified customers simultaneously without performance degradation.
AI-driven smart cart system that combines real-time data
acquisition, autonomous navigation, and predictive inventory III. SYSTEM ARCHITECTURE
management. The system leverages AI for decision-making The architecture consists of three primary layers—Hardware
and automation while employing ML models to learn from Layer, Data Processing Layer, and Application Layer—each
customer interactions, dynamically optimizing the shopping fulfilling specific functions to ensure optimal performance
process. By integrating RL for real-time navigation and and a seamless customer experience. An overview of these
LSTM networks for demand forecasting, our system adapts to layers is illustrated in Figure 1.
both the physical environment and shifting market conditions,
offering an end-to-end solution for retail automation. A. HARDWARE LAYER
The Hardware Layer is responsible for capturing real-time
C. AUTONOMOUS SYSTEMS FOR CUSTOMER data and enabling autonomous operations. It integrates
INTERACTION several essential components, including RGB-D cameras that
Autonomous systems, including robotic assistants and mobile provide both visual and depth data to support person tracking
service robots, are increasingly prevalent in retail and and obstacle detection. Ultrasonic sensors enhance obstacle
hospitality [21], [22]. These systems utilize AI for customer avoidance capabilities and help maintain safe distances from
guidance, real-time assistance, and autonomous naviga- customers, while infrared sensors detect line patterns or
tion [23]. For example, Simbe Robotics’ Tally [24] leverages predefined routes to aid in navigation. Product identification
AI-driven computer vision and ML to autonomously audit is automated through a barcode scanner or RFID reader,
store shelves, ensuring product availability. While similar which simplifies item addition as products are placed in the
technologies have been applied to autonomous shopping cart. Motors and controllers manage the cart’s movement
assistants, few integrate personalized customer interactions by adjusting speed and direction in response to sensor data.
with effective in-store navigation and inventory management. The onboard microcontroller (Raspberry Pi) processes these
Our smart shopping cart system enhances these autonomous inputs and executes low-latency commands for real-time
capabilities by incorporating personalized recommendations control, enabling smooth and responsive navigation.

55578 VOLUME 13, 2025

M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

where Kt is the Kalman gain, Pt represents the error

covariance, Ht is the observation matrix, and Rt denotes
measurement noise. For enhanced tracking, person re-
identification (Re-ID) employs cosine similarity to maintain
focus on the designated customer:
Fcust · Fref
S(Fcust , Fref ) = , (3)
∥Fcust ∥∥Fref ∥
where Fcust , representing the customer’s features, and Fref ,
a reference vector which may correspond to a previously
recorded or benchmark feature set. The term Fcust · Fref
represents the dot product of these two vectors, providing
a measure of how closely aligned they are in direction.
The quantities ∥Fcust ∥ and ∥Fref ∥ refer to the Euclidean
norms (or magnitudes) of the customer and reference vectors,
respectively, acting as scaling factors that normalize the
vectors.
Cloud processing supports computationally intensive
tasks, such as customer profiling, inventory management, and
data analytics. Customer profiling is conducted through deep
learning models that analyze past purchases and shopping
behavior, generating personalized product recommendations.
Inventory management is driven by an LSTM network that
forecasts demand to optimize stock levels. For demand
forecasting, the LSTM network updates its hidden states as
follows:
ht = LSTM(xt , ht−1 , ct−1 ), (4)
FIGURE 1. Overview of the AI-powered smart shopping cart.
where xt represents input data, ht is the hidden state, and
ct−1 is the cell state. Inventory prediction is modeled with
B. DATA PROCESSING LAYER the formula:
The Data Processing Layer serves as the backbone of the St+1 = St − Dt + Rt , (5)
system, executing AI models to drive decision-making. This
layer is divided into edge and cloud processing components where St+1 is the forecasted stock level, St the current stock,
to handle both low-latency tasks and more computationally Dt the predicted demand, and Rt the replenishment quantity.
intensive operations. Mean Squared Error (MSE) is minimized to optimize the
Edge processing, occurring directly on the cart’s onboard forecast:
n
microcontroller, ensures minimal latency for critical tasks 1X
like cart movement and obstacle avoidance, which are MSE = (Si − Spred,i )2 , (6)
n
i=1
essential for safe navigation. Person-specific tracking is
achieved through lightweight object detection models, where Si is the actual stock level and Spred,i the predicted
ensuring accurate customer tracking without overwhelming value.
computational resources. Additionally, path optimization is
facilitated by a RL model, enabling the cart to adapt to C. APPLICATION LAYER
dynamic environments and adjust its route based on store The Application Layer provides interfaces for customers and
layout and obstacles. For precise customer tracking, a Kalman administrators, allowing real-time interaction with the cart
filter is implemented. This filter tracks the customer’s and the backend system. The customer interface, accessible
position in real time with the following state update: via a mobile app or in-store display, offers features such as
personalized product recommendations based on shopping
x̂t+1 = Axt + But + wt , (1) data, shopping list integration that guides customers through
where xt is the current position, ut represents control input, the store, and automated billing with real-time item scanning
A is the state transition matrix, B is the control matrix, and for a seamless checkout experience. The administrator
wt accounts for process noise. The Kalman gain is iteratively interface enables store staff to monitor stock levels through
updated as follows: real-time updates and predictive analytics, observe sales
trends for optimizing store layout and promotions, and track
Kt = Pt HtT (Ht Pt HtT + Rt )−1 , (2) cart performance, including battery life and sensor status.

VOLUME 13, 2025 55579

 M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

IV. RESEARCH METHODS Algorithm 2 LSTM for Inventory Management

The proposed AI-powered smart shopping cart system 1: Input: Historical sales data X
is a comprehensive integration of hardware for real-time 2: Output: Forecasted demand D̂
data collection, ML models for decision-making, cloud 3: Initialize LSTM parameters Wf , Wi , Wo , bf , bi , bo
infrastructure for processing, and seamless POS system 4: for each time step t do
integration. 5: Update cell state and hidden state
6: end for
A. REINFORCEMENT LEARNING FOR PATH OPTIMIZATION 7: Return: Forecasted demand D̂
RL is used to optimize the cart’s navigation by dynamically
adjusting paths based on obstacles and store layout. The
Q-value is updated using the Bellman equation: secure data transmission, and Wi-Fi/Bluetooth for peripheral
connections. To protect customer data, AES-256 encryption
Q(st , at ) ← Q(st , at )
  and role-based access controls are implemented, ensuring
+ α rt + γ max Q(st+1 , a ) − Q(st , at ) , (7)
′ compliance with data security standards.
a′

where Q(st , at ) is the current Q-value, α is the learning rate, V. EXPERIMENTAL EVALUATION
rt represents the reward, γ is the discount factor, and a′ is the This section presents the experimental setup and performance
action that maximizes the next state Q-value. evaluation of the proposed system. Two datasets are utilized
to assess the system’s effectiveness in tracking, navigation,
B. DEEP LEARNING FOR CUSTOMER PERSONALIZATION recommendation, and demand forecasting, with comparison
Customer personalization is handled by a Convolutional against baseline algorithms across relevant metrics.
Neural Network (CNN) model, which processes customer
behavior data to generate tailored product recommendations. A. EXPERIMENTAL SETUP
The CNN operates with convolutional layers that identify The AI-powered smart shopping cart operates in a controlled
patterns in customer data, followed by max pooling and retail setting with multiple sensors and computing devices.
a fully connected layer to output recommendation scores. While these experiments provide valuable insights, future
Algorithm 1 outlines the CNN structure for generating these large-scale deployment in real retail stores will be necessary
recommendations. to assess system performance under varying real-world
conditions. The setup included the following components:
Algorithm 1 CNN for Customer Personalization • RGB-D Cameras: Capture visual and depth informa-
1: Input: Customer data X tion for customer tracking and obstacle detection.
2: Output: Product recommendations Ŷ • Ultrasonic and Infrared Sensors: Detect obstacles and
3: Initialize weights W and biases b maintain safe distances between the cart and customers.
4: for each customer i in X do • RFID Scanners and Barcode Readers: Automate
5: Perform convolution: Z ← W ∗ X [i] + b product identification as items are added to the cart.
6: Apply activation (ReLU): A ← max(0, Z ) • Onboard Computing (Raspberry Pi): Perform
7: Apply max pooling: P ← MaxPooling(A) real-time processing of sensor data and control cart
8: Flatten and apply fully connected layer: Ŷ ← σ (Wfc · movement.
F + bfc ) • Cloud Infrastructure: Handle computationally inten-
9: end for sive tasks, such as customer profiling, demand forecast-
10: Return: Ŷ ing, and generating personalized recommendations.

B. DATASETS
C. PREDICTIVE ANALYTICS FOR INVENTORY Two primary datasets are used for system evaluation:
MANAGEMENT • Retail Transaction Dataset (RTD): Contains historical
Inventory management leverages an LSTM-based predictive sales data from a large-scale retail environment and is
model to forecast demand. This model processes historical used for training the demand forecasting model, includ-
sales data to update the cell and hidden states and predict ing features such as product sales history, customer
future demand values, as shown in Algorithm 2. demographics, and purchase patterns.
• Shopping Mall Dataset (SMD): Consists of labeled
D. SYSTEM OPERATIONS AND SECURITY RGB-D video sequences from crowded environments,
The smart cart system integrates with POS systems to used to train and evaluate the customer tracking system.
enable real-time product scanning, automated billing, and This dataset provides a range of customer behaviors
contactless payment. Communication across the system and movement patterns, ideal for testing tracking
utilizes MQTT for low-latency updates, HTTP/HTTPS for robustness.

55580 VOLUME 13, 2025

M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

C. BASELINE ALGORITHMS accuracy, navigation efficiency, recommendation accuracy,

To validate the performance of the AI-powered smart demand forecasting, and system robustness.
shopping cart system, comparisons are made against the
following baseline algorithms commonly used in retail A. TRACKING ACCURACY
automation systems: We assess the tracking accuracy of the proposed system,
• A* Pathfinding Algorithm [26]: A standard algorithm which combines the Kalman filter with Re-ID, against
for finding the shortest path in static environments, used standalone Kalman filter and Re-ID approaches. Table 2
here to evaluate the cart’s navigation efficiency. and Figure 2 display the results, showing that the combined
• Dynamic Window Approach (DWA) [27]: A real-time method achieved a precision of 0.94, recall of 0.92, and
path planning algorithm that dynamically avoids obsta- F1-score of 0.93, surpassing the Kalman filter (F1-score:
cles, used to compare dynamic navigation capabilities in 0.83) and Re-ID alone (F1-score: 0.86). The ID Switch Rate
real-world environments. (IDSR) for the proposed method is significantly lower at 0.03,
• Collaborative Filtering [28]: A recommendation highlighting its robustness in crowded environments.
algorithm based on user-item interactions, used to We performed a statistical t-test to confirm that the preci-
evaluate recommendation accuracy. sion and F1-score improvements over standalone methods are
• ARIMA (Autoregressive Integrated Moving Aver- statistically significant (p < 0.05). The findings indicate that
age) [29]: A statistical model for time-series forecasting, the proposed system can reliably track individual customers
used as a baseline to compare the LSTM model’s in practical retail scenarios, improving user experience by
demand forecasting performance. reducing misidentifications and interruptions.
• Matrix Factorization (MF) [30]: A collaborative
filtering approach for generating personalized product
recommendations, providing a comparative perfor-
mance baseline with the system’s deep learning-based
recommendation model.

D. EVALUATION METRICS
The performance of the AI-powered smart shopping cart
system is evaluated using the following metrics:
• Tracking Accuracy: Assessed through precision,
recall, and F1-score, measuring the accuracy of
customer tracking and re-identification in crowded
environments.
• Navigation Efficiency: Metrics such as path length,
time to destination, collision rate, and success rate (task
completion without failure) are used to evaluate cart FIGURE 2. Tracking accuracy: F1-Score comparison.
navigation.
• Recommendation Accuracy: Measured using Preci-
sion@k, Recall@k, and Mean Reciprocal Rank (MRR) B. TRACKING ACCURACY AND ENVIRONMENTAL
to assess the relevance of recommended products based CONDITION ANALYSIS
on customer purchase history. To assess robustness, tracking accuracy is measured under
• Demand Forecasting Accuracy: Mean Squared Error different lighting (bright, dim, low-light) and crowd density
(MSE), Mean Absolute Error (MAE), and Root Mean conditions (low, medium, high). The Kalman filter + Re-ID
Squared Error (RMSE) are used to evaluate the LSTM method is compared with relevant baseline methods: Kalman
model’s demand forecasting. filter only, Re-ID model only, optical flow tracking, and
• User Satisfaction: Survey-based metrics such as YOLO-based object detection. As shown in Table 3, the
ease of use, recommendation relevance, and shopping Kalman filter + Re-ID approach maintained high accuracy
experience. in bright conditions (F1-score: 0.93) but showed slight
• Scalability and Robustness: Measured the system’s decreases in low-light (F1-score: 0.88) and high-density
ability to handle concurrent users, recover from failures, settings (F1-score: 0.90). This performance exceeded that of
and maintain real-time performance. the baseline methods under each condition, suggesting that
the combined Kalman filter + Re-ID model is effective under
VI. RESULTS standard conditions and moderately robust in challenging
This section presents the evaluation results of the proposed environments.
AI-powered smart shopping cart system, compared with The tracking accuracy of the baseline methods generally
baseline algorithms across various aspects such as tracking decreased under low-light and high-density conditions, with

VOLUME 13, 2025 55581

 M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

TABLE 2. Tracking accuracy results using SMD.

optical flow and YOLO-based detection particularly sensitive demonstrating its ability to provide relevant and personalized
to crowd density and lighting variations. These results recommendations.
underscore the advantages of combining Kalman filtering and
Re-ID for robust tracking in retail environments. G. REAL-TIME DEMAND FORECASTING IN DYNAMIC
RETAIL SCENARIOS
C. TRACKING ROBUSTNESS To evaluate adaptability, the LSTM-based demand forecast-
To assess the system’s robustness, tracking accuracy is tested ing model is tested in scenarios with simulated demand
under varying conditions such as high crowd density and surges (e.g., flash sales or promotions). Table 6 shows the
low lighting. In high-density settings, the tracking system model’s accuracy over time, with Mean Squared Error (MSE)
maintained an F1-score of 0.90 with a minor increase in the increasing slightly during the surge but stabilizing afterward.
ID Switch Rate to 0.05. However, in low-light conditions, The LSTM model adjusted within 5 minutes of the demand
performance dropped slightly with an F1-score of 0.88, spike, demonstrating its responsiveness to rapid changes.
indicating sensitivity to lighting conditions. This analysis
suggests that adding thermal or infrared sensors could H. DEMAND FORECASTING ACCURACY
improve robustness in adverse conditions. The LSTM-based demand forecasting model demonstrated
superior accuracy over ARIMA, achieving lower MSE
D. NAVIGATION EFFICIENCY AND REAL-TIME COLLISION (0.034), MAE (0.055), and RMSE (0.046) as presented in
AVOIDANCE TESTING WITH DYNAMIC OBSTACLES Table 7 and Figure 5. The forecast bias is also significantly
The system’s navigation efficiency is evaluated in environ- lower, indicating the model’s capability to manage inventory
ments with static and dynamic obstacles, specifically testing effectively and reduce stockouts.
real-time collision avoidance. The RL-based navigation
system is compared with two established baseline methods: I. USER SATISFACTION ANALYSIS
A* pathfinding and Dynamic Window Approach (DWA). User surveys are conducted to evaluate satisfaction with
As shown in Table 4 and Figure 3, the RL-based system system features, such as ease of use, recommendation
maintained a collision rate below 0.03 and achieved a success relevance, and overall shopping experience. Results indicated
rate of 96% in high-density and dynamic obstacle settings. an 85% satisfaction rate, with high ratings in ease of use and
The RL-based system outperformed both A* and DWA, recommendation quality. User feedback also highlighted the
especially in dynamic obstacle scenarios, where it maintained convenience of automated billing and the accuracy of real-
lower collision rates and shorter path lengths, showcasing its time recommendations.
adaptive capability to avoid obstacles in real-time.
J. SCALABILITY AND SYSTEM LATENCY
E. NAVIGATION RELIABILITY AND ROBUSTNESS The system is tested with up to 50 concurrent carts in
The navigation system is further evaluated under condi- a simulated retail environment. Average system latency
tions with varying obstacle density and dynamic obstacles remained below 200 ms, ensuring real-time responsiveness.
(e.g., moving individuals). The RL-based navigation system Performance scaled effectively, with minor latency increases
showed a 96% success rate in high-density obstacle scenarios, observed as the number of carts approached capacity,
compared to 85% and 89% for A* and DWA, respectively. indicating robustness under increased loads.
The RL approach dynamically adjusted paths in response to
moving obstacles, demonstrating adaptability that is critical K. SUMMARY OF RESULTS
in retail environments with high foot traffic. The proposed AI-powered smart shopping cart system
consistently outperformed baseline algorithms across all
F. RECOMMENDATION ACCURACY evaluated tasks:
The deep learning-based recommendation system outper- • Tracking Accuracy: The combined Kalman filter + Re-
formed collaborative filtering (CF) and matrix factorization ID approach achieves higher precision, recall, and F1-
(MF) in terms of Precision@5, Recall@5, and Mean score, with a significantly reduced ID Switch Rate.
Reciprocal Rank (MRR). As shown in Table 5 and Figure 4, • Navigation Efficiency and Robustness: RL-based
the proposed method reached a Precision@5 of 0.88, higher navigation demonstrated shorter paths, reduced travel
than CF (0.76) and MF (0.80), with an MRR of 0.84, time, and higher adaptability in dynamic environments.

55582 VOLUME 13, 2025

M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

TABLE 3. Tracking accuracy across environmental conditions for proposed and baseline methods using SMD.

TABLE 4. Navigation efficiency and collision avoidance in dynamic environments using the SMD.

FIGURE 3. Collision avoidance in dynamic environments.

TABLE 5. Recommendation system performance (k=5) using RTD. TABLE 7. Demand forecasting accuracy results using RTD.

• Demand Forecasting Accuracy: The LSTM model

TABLE 6. Demand forecasting accuracy during normal and surge events
using RTD. demonstrated higher accuracy with reduced forecasting
errors and bias, supporting effective inventory manage-
ment.
• User Satisfaction and Scalability: High user satisfac-
tion rates and robust scalability indicate the system’s
potential for practical deployment.
• Recommendation Accuracy: The deep learning model These results validate the system’s capability to enhance
provided more relevant and diverse recommendations retail operations by improving tracking, navigation, person-
compared to CF and MF. alization, and demand forecasting.

VOLUME 13, 2025 55583

 M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

deployment in retail stores may introduce additional chal-

lenges. Factors such as varying lighting conditions, high
customer density, and complex store layouts could influence
tracking accuracy and navigation efficiency. Future work will
focus on large-scale testing in diverse retail settings to assess
these impacts. Additionally, integrating adaptive AI models
that can dynamically adjust to environmental conditions will
further enhance system robustness and usability in real-world
scenarios.

ACKNOWLEDGMENT
(Muhammad Imran Zulfiqar and Ayesha Khalid contributed
equally to this work.)

REFERENCES
FIGURE 4. Recommendation accuracy: Precision@5 comparison. [1] H.-P. Lu, H.-L. Cheng, J.-C. Tzou, and C.-S. Chen, ‘‘Technology roadmap
of AI applications in the retail industry,’’ Technol. Forecasting Social
Change, vol. 195, Oct. 2023, Art. no. 122778.
[2] M. R. Kumar, J. Venkatesh, and A. M. J. M. Z. Rahman, ‘‘Data mining
and machine learning in retail business: Developing efficiencies for better
customer retention,’’ J. Ambient Intell. Humanized Comput., vol. 2021,
pp. 1–13, Jan. 2021.
[3] M. Sharma, A. Tomar, and A. Hazra, ‘‘Edge computing for industry 5.0:
Fundamental, applications and research challenges,’’ IEEE Internet Things
J., 2024.
[4] I. Ahmed, G. Jeon, and F. Piccialli, ‘‘From artificial intelligence to
explainable artificial intelligence in industry 4.0: A survey on what, how,
and where,’’ IEEE Trans. Ind. Informat., vol. 18, no. 8, pp. 5031–5042,
Aug. 2022.
[5] R. Wolniak, K. Stecuła, and B. Aydın, ‘‘Digital transformation of grocery
in-store shopping-scanners, artificial intelligence, augmented reality and
beyond: A review,’’ Foods, vol. 13, no. 18, p. 2948, Sep. 2024.
[6] L. A. Escudero-Gómez, ‘‘Shopping centers challenging decline: Competi-
tive strategies in three case studies from Madrid’s urban area,’’ J. Retailing
Consum. Services, vol. 79, Jul. 2024, Art. no. 103826.
[7] R. Narang and S. Tiwari, ‘‘Role of modern technology in unorganized retail
sector,’’ J. Knowl. Econ., vol. 2024, pp. 1–28, May 2024.
[8] A. Ala, A. H. Sadeghi, M. Deveci, and D. Pamucar, ‘‘Improving smart deals
FIGURE 5. Demand forecasting accuracy: MSE comparison.
system to secure human-centric consumer applications: Internet of Things
and Markov logic network approaches,’’ Electron. Commerce Res., vol. 24,
no. 2, pp. 771–797, Jun. 2024.
[9] S. R. Sandeep, S. Ahamad, D. Saxena, K. Srivastava, S. Jaiswal, and
VII. CONCLUSION
A. Bora, ‘‘To understand the relationship between machine learning and
In conclusion, the proposed AI-powered smart shopping artificial intelligence in large and diversified business organisations,’’
cart system successfully demonstrates the transformative Mater. Today, Proc., vol. 56, pp. 2082–2086, Jan. 2022.
[10] A. H. Sodhro, S. Pirbhulal, and V. H. C. de Albuquerque, ‘‘Artificial
potential of advanced AI and ML technologies in the retail intelligence-driven mechanism for edge computing-based industrial appli-
sector. By integrating person-specific tracking, RL for path cations,’’ IEEE Trans. Ind. Informat., vol. 15, no. 7, pp. 4235–4243,
optimization, and LSTM-based demand forecasting, the Jul. 2019.
[11] Y. K. Dwivedi et al., ‘‘Artificial intelligence (AI): Multidisciplinary
system enhances both the shopping experience and oper- perspectives on emerging challenges, opportunities, and agenda for
ational efficiency. Through real-time data collection, edge research, practice and policy,’’ Int. J. Inf. Manage., vol. 57, Aug. 2019,
computing, and cloud-based decision-making, this system Art. no. 101994.
[12] S. Karjol, A. K. Holla, C. B. Abhilash, P. V. Amrutha, and Y. V. Manohar,
addresses key challenges in modern retail, such as inefficient ‘‘An IoT based smart shopping cart for smart shopping,’’ in Proc. Int. Conf.
inventory management, suboptimal customer experiences, Cognit. Comput. Inf. Process., Jan. 2018, pp. 373–385.
and manual checkout processes. The experimental results [13] M. A. Sarwar, Y. A. Daraghmi, K. W. Liu, H. C. Chi, T. U. Ik, and Y. L. Li,
‘‘Smart shopping carts based on mobile computing and deep learning
show significant improvements in customer satisfaction,
cloud services,’’ in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC),
operational efficiency, and inventory management when May 2020, pp. 1–6.
compared to baseline systems. This system provides a fully [14] M. Schmitt, ‘‘Automated machine learning: AI-driven decision making in
autonomous and adaptive retail solution that not only reduces business analytics,’’ Intell. Syst. Appl., vol. 18, May 2023, Art. no. 200188.
[15] D. Grewal, S. Benoit, S. M. Noble, A. Guha, C.-P. Ahlbom, and
costs but also enhances the overall shopping experience J. Nordfält, ‘‘Leveraging in-store technology and AI: Increasing customer
through personalized recommendations and seamless billing and employee efficiency and enhancing their experiences,’’ J. Retailing,
integration. vol. 99, no. 4, pp. 487–504, Dec. 2023.
[16] C. Campbell, S. Sands, C. Ferraro, H.-Y.-J. Tsao, and A. Mavrommatis,
While our system has demonstrated strong performance ‘‘From data to action: How marketers can leverage AI,’’ Bus. Horizons,
in controlled environments, we acknowledge that real-world vol. 63, no. 2, pp. 227–243, Mar. 2020.

55584 VOLUME 13, 2025

M. I. Zulfiqarx et al.: AI-Driven Smart Shopping Carts With Real-Time Tracking and Inventory Forecasting

[17] K. Yan, L. Liu, Y. Xiang, and Q. Jin, ‘‘Guest editorial: AI and machine AYESHA KHALID received the B.S.C.S. degree
learning solution cyber intelligence technologies: New methodologies and from the Department of Computer Science and IT,
applications,’’ IEEE Trans. Ind. Informat., vol. 16, no. 10, pp. 6626–6631, University of Jhang, Jhang, Pakistan, in 2024.
Oct. 2020. She has research focused on the Internet of
[18] A. M. Pereira, J. A. B. Moura, E. D. B. Costa, T. Vieira, Medical Things for Hand Gesture Recognition
A. R. D. B. Landim, E. Bazaki, and V. Wanick, ‘‘Customer models Systems, which lies at the intersection of machine
for artificial intelligence-based decision support in fashion online retail learning and deep learning.
supply chains,’’ Decis. Support Syst., vol. 158, Jul. 2022, Art. no. 113795.
[19] M.-A. Blais and M. A. Akhloufi, ‘‘Reinforcement learning for swarm
robotics: An overview of applications, algorithms and simulators,’’ Cognit.
Robot., vol. 3, pp. 226–256, Jan. 2023.
[20] M. I. Zulfiqar, I. Abdullah, A. Akram, S. Mehmood, Fareeha, and A. Ali,
‘‘An efficient autonomous maze solving robot based on microcontroller by
avoiding obstacles using a decision-making algorithm,’’ in Proc. Int. Conf.
IT Ind. Technol. (ICIT), Oct. 2023, pp. 1–6.
[21] J. Reis, N. Melão, J. Salvadorinho, B. Soares, and A. Rosete, ‘‘Service
robots in the hospitality industry: The case of henn-na hotel, Japan,’’
Technol. Soc., vol. 63, Nov. 2020, Art. no. 101423. ABUBAKR SIDDIG received the Ph.D. degree
[22] S. Mukherjee, M. M. Baral, R. Nagariya, C. Venkataiah, U. V. A. Rao, and in Computer Science. He is currently a Faculty
K. S. Rao, ‘‘Systematic literature review and future research directions for Member with the School of Computer Science,
service robots in hospitality and tourism industries,’’ Service Industries J., Technological University Dublin, Ireland. He has
vol. 43, nos. 15–16, pp. 1083–1116, Dec. 2023. contributed extensively to the fields of localization
[23] H. He, J. Gray, A. Cangelosi, Q. Meng, T. M. McGinnity, and J. Mehnen, and AI-driven applications, with publications in
‘‘The challenges and opportunities of human-centered AI for trustworthy reputable international journals and conferences.
robots and autonomous systems,’’ IEEE Trans. Cognit. Develop. Syst., He is actively involved in interdisciplinary projects
vol. 14, no. 4, pp. 1398–1412, Dec. 2022.
aimed at advancing computational methods. His
[24] D. Mronga, A. Bresser, F. Maas genannt Bermpohl, A. Danzglock,
research interests include data science, natural
S. Stelter, A. Hawkin, H. G. Nguyen, M. Beetz, and F. Kirchner,
‘‘MARLIN: A cloud integrated robotic solution to support intralogistics language processing, data mining, and speech and audio quality analysis.
in retail,’’ Robot. Auto. Syst., vol. 175, May 2024, Art. no. 104654.
[25] I. Keshta, ‘‘AI-driven IoT for smart health care: Security and privacy
issues,’’ Informat. Med. Unlocked, vol. 30, Jan. 2022, Art. no. 100903.
[26] F. Grenouilleau, W.-J.-V. Hoeve, and J. N. Hooker, ‘‘A multi-label A
algorithm for multi-agent pathfinding,’’ in Proc. Int. Conf. Automated
Planning Scheduling, vol. 29, pp. 181–185, Jul. 2019.
[27] M. Kobayashi and N. Motoi, ‘‘Local path planning: Dynamic window
approach with virtual manipulators considering dynamic obstacles,’’ IEEE MUHAMMAD JUNAID NAWAZ received the
Access, vol. 10, pp. 17018–17029, 2022. B.S. and M.S. degrees in Computer Science from
[28] M. K. Najafabadi, M. N. Mahrin, S. Chuprat, and H. M. Sarkan, the Qurtuba University of Science and Infor-
‘‘Improving the accuracy of collaborative filtering recommendations using mation Technology, Dera Ismail Khan, Khyber
clustering and association rules mining on implicit data,’’ Comput. Hum. Pakhtunkhwa, Pakistan.
Behav., vol. 67, pp. 113–128, Feb. 2017. He has research focused on the Internet of
[29] R. H. Shumway, D. S. Stoffer, R. H. Shumway, and D. S. Stoffer, ‘‘ARIMA Medical Things for Hand Gesture Recognition
models,’’ in Time Series Analysis and Its Applications: With R Examples. Systems, which lies at the intersection of machine
USA: Springer, 2017, pp. 75–163. learning and deep learning.
[30] D. Bokde, S. Girase, and D. Mukhopadhyay, ‘‘Matrix factorization model
in collaborative filtering algorithms: A survey,’’ Proc. Comput. Sci.,
vol. 49, pp. 136–146, Jan. 2015.

MUHAMMAD IMRAN ZULFIQAR received

the B.Sc. degree in Computer Engineering from SALIM SAAY received the Ph.D. degree in
the University of Engineering and Technology, Computer Science from Kabul University.
Lahore, Pakistan, the M.S. degree in Com- He is currently an Associate Professor with the
puter Science from the University of Agriculture Department of Computer Science and Information
Faisalabad, Pakistan, and the Ph.D. degree from Systems, University of Limerick, Ireland, and a
the School of Computer Science and Engineering, member of Lero—the SFI Research Centre for
Nanjing University of Science and Technology, Software. He has held academic and research
Nanjing, China. positions with the Technological University of
He has published some technical papers in Shannon (TUS) and Kabul University. He has
prominent journals and conferences, such as IEEE INTERNET OF THINGS authored several research papers in prestigious
JOURNAL, IEEE ACCESS, Fundamental Research, and PeerJ Computer journals and conferences, including IEEE TRANSACTIONS and other prominent
Science. His research interests include machine learning, data mining, the venues. His research interests include software architecture, blockchain inte-
Internet of Things, information security, and robotics. He also serves as gration, artificial intelligence, e-learning, and organizational collaboration
a Reviewer for IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING technologies. He also serves as a reviewer for top-tier journals.
SYSTEMS and Pattern Recognition Letters.

VOLUME 13, 2025 55585