ARTIFICIAL INTELLIGENCE BASED

SHOPPING CART DESIGNED ON NEURAL

NETWORK AND MOBILENETv2
FRAMEWORK
Bindu S Reddy Bhargavi P Ayush Roy
Electronics and Communication Electronics and Communication Electronics and Communication
Engineering Engineering Engineering
Cambridge Institute of Technology, Cambridge Institute of Technology, Cambridge Institute of Technology,
Bengaluru, INDIA Bengaluru, INDIA Bengaluru, INDIA
bindums077@[Link] bhargaviprasanna023@[Link] Aayushrai0023@[Link]

Chandra Shekar B Indu K

Electronics and Communication Electronics and Communication
Engineering Engineering
Cambridge Institute of Technology,
Cambridge Institute of Technology,
Bengaluru, INDIA
Bengaluru, INDIA
[Link]@[Link]
Chandru812321@[Link]

Abstract— This paper presents an AI-powered self- The motivation behind this research stems from the need to
checkout system, iCheckout, designed to improve the retail enhance speed, accuracy, and affordability in automated
shopping experience by integrating machine learning, retail systems. Current cashier-less technologies, such as
computer vision, and edge computing. The system employs Amazon Go, rely heavily on complex infrastructure, making
MobileNetV2 for product identification and an HX711-based
them expensive and challenging to scale. iCheckout, on the
load cell weight sensor for verification, ensuring high accuracy
and efficiency. Unlike traditional barcode-based checkouts, other hand, is designed to be lightweight and efficient by
iCheckout eliminates the need for human intervention, offering utilizing MobileNetV2 for image-based product
a seamless and contactless transaction process. Experimental identification and an HX711-based load cell for weight
results demonstrate a 90% object detection accuracy and a validation. This hybrid approach ensures high accuracy
74% precision rate for weight verification, with optimized edge while minimizing computational overhead, making it ideal
processing to reduce latency. The system is designed for for real-world retail environments.
scalability and cost-effectiveness, making it a practical solution
for smart retail automation.
II. LITERATURE REVIEW
Keywords— Autonomous checkout, AI in retail, Edge
computing, Computer vision, MobileNetV2 Kumar et al. [1] introduced an AI-powered automated
billing system using Raspberry Pi, demonstrating its
I. INTRODUCTION feasibility for low-cost embedded applications. However,
the study highlighted computational limitations that affected
The growing demand for automation in retail has led to the real-time processing and constrained the complexity of AI
development of various self-checkout solutions aimed at models deployed on edge devices.
reducing customer wait times and improving transaction
efficiency. Traditional barcode scanning systems often lead to Lee et al. [2] proposed a deep learning-based cashier-less
long queues, human errors, and operational inefficiencies, checkout system leveraging Convolutional Neural Networks
necessitating the need for an advanced, AI-driven alternative. (CNNs) for product recognition. Their work improved
iCheckout addresses these challenges by integrating AI- classification accuracy, but high computational complexity
powered object recognition and real-time weight-based and power consumption posed scalability challenges,
verification to create an intelligent self-checkout solution. particularly for large-scale retail environments.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

To address latency issues in AI-driven retail automation, Wilson et al. [12] examined the integration of blockchain
Palacios et al. [3] explored multi-sensor synchronization for and IoT to establish a secure transaction framework for
improved product recognition. By integrating multiple autonomous checkout systems. Their research emphasized
sensors, they optimized data flow, reducing processing the role of decentralized ledgers in ensuring data integrity,
delays and enhancing real-time accuracy. fraud prevention, and tamper-proof payment processing,
highlighting the potential of blockchain-driven retail
Tan et al. [4] examined an enhanced self-checkout system solutions.
using the YOLOv10 object detection model, which
significantly improved recognition speed and accuracy. De Bellis et al. [13] focused on consumer adoption
However, the computational demand of YOLOv10 raised challenges, emphasizing the importance of user-friendly
concerns regarding its feasibility for deployment on system designs to enhance trust, engagement, and
resource-constrained embedded devices, underscoring the accessibility in AI-driven retail environments. Their study
need for lightweight AI models. underscored the need for intuitive interfaces and reliable AI
models to improve customer experience and increase
Sauter and Treytl [5] analyzed security and privacy risks in adoption rates.
IoT-enabled checkout systems. Their research identified
vulnerabilities related to authentication and operational Garcia et al. [14] explored IoT-enabled real-time billing
security, emphasizing the necessity of robust data protection systems, demonstrating how seamless connectivity and
mechanisms to prevent unauthorized access. automated transaction processing could significantly reduce
checkout time. Their work highlighted the potential for
Falcão et al. [6] developed Isaacs, an intra-store cloud-synced retail platforms, allowing real-time price
autonomous checkout system optimized for low-power adjustments and inventory updates, which contribute to
embedded platforms. Their study demonstrated that greater operational efficiency in automated stores.
lightweight AI models could balance computational
efficiency and accuracy, ensuring real-time product Trigo et al. [15] investigated scalable AI-driven checkout
recognition with minimal resource consumption. solutions for high-traffic environments, addressing
performance limitations through optimized AI models and
Biglari and Tang [7] explored embedded machine learning distributed computing architectures. Their research
techniques, identifying MobileNetV2 as an effective provided valuable insights into enhancing the speed,
architecture for balancing processing efficiency and reliability, and scalability of large-scale autonomous retail
classification accuracy. Their findings reinforce the systems.
suitability of optimized AI models for autonomous checkout
applications. AiFi Inc. [16] analyzed market trends in cashier-free
retail, reinforcing the growing consumer preference for AI-
Wen et al. [8] investigated the role of edge computing in powered self-checkout solutions. Their study highlighted
retail automation, showcasing its ability to reduce latency how advancements in computer vision, deep learning, and
and improve system responsiveness. Their work IoT integration are driving demand for scalable, automated
demonstrated that leveraging edge AI for real-time product shopping experiences.
recognition minimizes reliance on cloud infrastructure,
enhancing transaction speed and security. Szabó-Szentgróti et al. [17] provided a systematic review of
cashier-less store technologies, validating the feasibility of
Gupta et al.[9] explored contactless payments in IoT real-time deep learning models for product recognition and
ecosystems, emphasizing the importance of seamless and transaction processing. Their research demonstrated that AI-
secure transactions in AI-driven checkout solutions. Their powered checkout systems could significantly reduce
findings align with the growing trend of blockchain operational costs while enhancing efficiency in smart
integration to enhance security and fraud prevention in retail environments.
automated retail.
Huang et al. [18] investigated edge AI techniques for
Wilson et al. [10] further examined the convergence of optimizing real-time processing in checkout automation. Their
blockchain and IoT in autonomous checkout systems, study emphasized the role of low-latency inference models in
proposing a secure transaction framework that ensures data reducing dependency on cloud-based computation, ensuring
integrity and minimizes risks associated with digital faster and more reliable product recognition.
payments.
Rossi et al. [19] explored customer behavior in
Fouladi et al. [11] introduced hybrid product detection autonomous stores, emphasizing the importance of fast
systems, integrating multiple sensing technologies such as and efficient checkout experiences to maintain user
vision and weight-based validation to enhance accuracy. satisfaction. Their study reinforced the need for low-latency
Their findings demonstrated that combining sensor fusion AI models and seamless payment integration to enhance
techniques minimizes misclassification errors and improves the shopping experience.
the reliability of product identification.
Shin et al. [20] proposed advancements in object detection 2. Object Classification Module: MobileNetV2
and payment integration, addressing the technical extracts key product features, such as shape,
challenges of real-time retail automation. Their research texture, and color, ensuring high accuracy in
contributed to the development of optimized AI models for recognition.
product recognition, ensuring high-speed and error-free 3. Weight Verification Module: An HX711 load
checkout transactions. Building on these studies, this cell sensor measures the weight of the identified
research implements iCheckout, a hybrid AI-based product, validating it against stored reference
checkout system that integrates MobileNetV2 for object values to prevent errors.
detection, weight sensors for validation, and edge 4. Transaction and Data Management Module: A
computing for low-latency processing. Flask-based web interface displays checkout
details in real-time, while a database management
III. MATERIALS AND METHODS system (MySQL) stores product details, pricing,
This project follows a structured and systematic approach to and past transactions for seamless processing.
designing, implementing, and validating an AI-driven
autonomous checkout system, iCheckout. The methodology A state-machine-based approach is employed to optimize
ensures the integration of computer vision, deep learning, data flow, minimize latency, and enhance computational
weight-based validation, and edge computing while efficiency. The system architecture is modular, allowing
optimizing transaction speed, accuracy, and security. The scalability for integration into large-scale retail
development process is divided into four key stages: environments. Edge computing principles are applied to
literature review, system design, implementation, and process AI inference locally on Raspberry Pi, ensuring fast
simulation/testing. Each stage is crucial in ensuring a and secure transactions. The final design is optimized to
robust, scalable, and efficient AI-powered self-checkout maintain low power consumption, high-speed processing,
system for retail automation. and robust error handling for real-world deployment.

C. Implementation
A. Literature Review
The implementation of iCheckout involves both hardware
and software components to ensure efficient real-time
The research begins with an extensive analysis of AI-driven product classification, validation, and billing. The system
retail automation technologies, focusing on computer is developed using Python, TensorFlow Lite, OpenCV,
vision, machine learning, and edge computing. A review and Flask, with hardware integration on Raspberry Pi
of existing cashier-less checkout solutions provides for edge computing.
insights into their advantages and limitations, particularly in
product recognition, transaction security, and system
1. Software Implementation
scalability. Special emphasis is placed on MobileNetV2, a
lightweight CNN model optimized for real-time object
The software development process involves coding and
classification in embedded systems. The literature also
optimizing AI-based object detection, weight-based
explores sensor-based verification techniques, including
verification, and transaction processing. The
the use of HX711 load cell sensors for weight-based
implementation includes:
authentication to prevent misclassification errors.
1.1 Image Processing and Object Classification
Additionally, studies on edge computing for real-time AI
inference are analyzed to optimize latency reduction and
The computer vision module is responsible for real-time
computational efficiency in embedded platforms like
product recognition using OpenCV and MobileNetV2,
Raspberry Pi. The review also covers security frameworks
ensuring accurate classification with minimal
for AI-driven transactions, including blockchain-based
computational overhead.
tamper-proof payment processing and IoT-enabled fraud
detection mechanisms. These insights serve as a foundation
for developing iCheckout, ensuring high accuracy and • Feature Extraction: MobileNetV2 analyzes key
efficiency while addressing key challenges in real-time product features such as shape, color, and texture
retail automation. for precise identification.
• Transfer Learning: The model is fine-tuned on a
B. System Design custom dataset, improving accuracy for diverse
product categories.
The iCheckout system is designed to enable autonomous, • Real-Time Processing: The system captures and
fast, and error-free product identification and billing. The classifies images in under one second, ensuring a
architecture consists of four primary modules: seamless checkout experience.
• Lighting and Angle Adaptation: Preprocessing
1. Image Processing Module: A high-resolution techniques enhance image quality, reducing errors
camera module captures product images, which caused by poor lighting or occlusions.
are preprocessed and sent to the MobileNetV2 AI • Error Handling: If an item is misclassified, the
model for classification. system re-runs classification or requests manual
verification to prevent billing errors.
1.2 Weight-Based Verification

The HX711 load cell sensor ensures fraud prevention and

transaction accuracy by verifying the product’s weight
against stored reference values.

• Real-Time Measurement: The sensor continuously

records precise weight data, ensuring reliable
verification.
• Fraud Prevention: If an object’s weight does not
match its expected value, the system flags a
potential mismatch and prompts corrective action.
• Automatic Calibration: The system periodically
adjusts for environmental factors (e.g.,
temperature, sensor drift) to maintain accuracy.
• Multi-Item Detection: If multiple items are placed
at once, the system can identify inconsistencies
Figure 1: Process Flow of Automated Checkout System
and request separation.
• Adaptive Thresholding: Small weight variations 1.4 Edge Computing Optimization
are accounted for using tolerance limits, reducing
false alarms while maintaining security.
To minimize latency and enhance performance,
iCheckout leverages edge computing on Raspberry Pi,
reducing reliance on cloud-based processing.
1. .3 Database Management and Web Interface
• Local AI Inference: TensorFlow Lite allows AI
models to run directly on the Raspberry Pi,
A Flask-based user interface enables seamless
eliminating the need for cloud servers.
transaction processing, ensuring real-time display and
secure data storage. • Latency Reduction: Edge computing ensures that
transactions are processed in milliseconds,
preventing checkout delays.
• User-Friendly Interface: Customers can view
• Energy Efficiency: Optimized deep learning
product details, weight verification, and billing
models consume low power, making the system
information in real time.
suitable for portable or battery-operated devices.
• Secure Data Storage: A MySQL database
maintains product details, weight thresholds, and • AI Model Compression: Quantization and
pruning techniques reduce model size without
transaction history, ensuring data integrity.
compromising accuracy, improving performance
• Fast Retrieval: Optimized database queries allow on embedded devices.
instant product lookup, reducing processing
delays. • High-Traffic Readiness: The system maintains
stable performance even under high transaction
• Scalability: The database supports large-scale loads, ensuring scalability for busy retail
retail operations, handling thousands of
environments.
transactions efficiently.
• Data Integrity: The system prevents duplicate
entries and data corruption, ensuring accurate
billing and transaction history.

Figure 2: System Architecture for Automated Checkout

D. Simulation and Testing

The implemented iCheckout system undergoes

comprehensive simulation, validation, and real-world
testing to ensure optimal performance in a retail different product types. Some product categories had fewer
environment. The testing phase is divided into software training samples, but augmentation techniques helped
simulations and hardware-based evaluations to assess the mitigate any bias. The evaluation metrics, including
system’s accuracy, efficiency, and robustness. precision, recall, and F1-score, confirmed the reliability of
the trained model.
Initially, MATLAB simulations are conducted to analyze
the classification accuracy of the MobileNetV2 model.
Various datasets of retail products are used to fine-tune the
model, ensuring it correctly identifies and differentiates
items based on real-time image inputs. The confusion
matrix, precision, recall, and F1-score are evaluated to
measure model performance.

For hardware-level validation, the Raspberry Pi-based

implementation undergoes benchmarking to assess weight
measurement precision using the HX711 sensor module.
Multiple test cases with different product weights are Figure 3: Simulated Waveform of MobileNetV2 Training
analyzed to compare measured values against reference Model
weights, ensuring minimal error. The system’s latency is
B. Real-Time Hardware Results (Hardware Approach)
also tested to determine inference speed, transaction
processing time, and overall responsiveness, ensuring After successful software simulations, the trained model
seamless operation even in high-traffic retail environments. was deployed on Raspberry Pi 4 for real-world testing.
This phase introduced additional challenges such as varying
Additionally, real-world testing is conducted in simulated lighting conditions, object positioning, and processing
store conditions, where users interact with the system for power limitations. Despite these factors, the system
product scanning, validation, and checkout. The maintained a classification accuracy of 94.2%, slightly
transaction efficiency is evaluated by measuring the speed lower than the simulated results due to real-world
of item classification, weight verification, and payment complexities. The HX711 load cell sensor was integrated
processing through the Flask-based web interface. Errors, for weight verification, achieving an average weight
discrepancies, and failure cases are recorded to refine precision of 74%, with some minor deviations when
system performance, ensuring a reliable, fast, and objects were placed incorrectly or off-center on the sensor.
automated checkout experience. The image classification time on Raspberry Pi was
measured at 750ms per image, which was significantly
higher than the 250ms processing time recorded in
IV. RESULTS AND DISCUSSIONS
MATLAB simulations due to the limited computational
The iCheckout system was evaluated using both simulation resources of embedded systems. The confidence scores
and hardware implementations to assess its performance generated by the model also varied, with controlled
in real-world scenarios. The evaluation focused on object simulations yielding 85-100% confidence levels, while
detection accuracy, weight measurement precision, real-world testing resulted in fluctuations between 75-
computational efficiency, and transaction speed. 100%, requiring additional verification steps in some cases.
MATLAB simulations were conducted to optimize the
MobileNetV2 deep learning model, while real-time testing
on Raspberry Pi provided insights into system limitations
and practical challenges. The results indicate that the system
performs efficiently under controlled conditions but faces
slight accuracy drops when deployed in real-world
environments due to factors like lighting inconsistencies,
occlusions, and sensor drift.
A. Simulation Results (Software Approach)
MATLAB simulations were used to optimize and validate
the MobileNetV2-based object detection model before
deployment. The model was trained and tested on a dataset Figure 4: Predicted Class and Confidence Score
consisting of six product categories, achieving a
classification accuracy of 98.5% under controlled C. Comparison Between Simulation and Hardware
conditions. The training performance analysis showed a Implementation
smooth convergence of loss values, indicating effective A comparative analysis was conducted to highlight the
learning of product features. Various dataset augmentation differences between MATLAB-based software simulations
techniques, such as flipping, contrast adjustments, and noise and real-time hardware implementation on Raspberry Pi.
addition, were employed to enhance the model's robustness. The simulation phase demonstrated high accuracy and fast
The impact of class distribution on accuracy was also inference times due to the computational power of MATLAB,
analyzed, ensuring that the model generalized well across while the hardware deployment faced processing
limitations and environmental factors that slightly affected [5] Sauter, T., and Treytl, A., "IoT-Enabled Sensors in
accuracy and efficiency. The classification accuracy Automation Systems and Their Security Challenges," in
dropped from 98.5% in MATLAB to 94.2% in Raspberry IBEE Sensors Letters, vol. 7, no. 12, pp. 1-4, Dec. 2023, Art
Pi deployment, mainly due to lighting inconsistencies and no. 7500904, doi: 10.1109/LSENS.2023.3332404.
occlusions. Similarly, the processing speed decreased from [6] Falcão, J. D., Ruiz, C., Bannis, A., Noh, H. Y., and
250ms per image in MATLAB to 750ms per image in Zhang, P., "Isacs: In-store autonomous checkout system for
Raspberry Pi, highlighting the impact of hardware constraints. retail," Proceedings of the ACM on Interactive, Mobile,
The confidence score remained stable in MATLAB but Wearable and Ubiquitous Technologies, vol. 5, no. 3, pp. 1-
fluctuated in real-world conditions, necessitating an adaptive 26, 2021.
verification mechanism. Overall, while the software [7] Biglari, Amin, and Tang, Wei, "A Review of Embedded
implementation provided near-perfect results, the hardware Machine Learning Based on Hardware, Application, and
deployment required additional optimizations to enhance Sensing Scheme," Sensors, vol. 23, no. 4, p. 2131, 2023.
reliability in practical retail settings. [Link]
[8] Vishnoi, V. K., Ramesh, P., Choudhary, A., and Patel,
D., "Detection of Diseases in Apple Plants Using CNN,"
International Journal of Artificial Intelligence, 2022.
[9] Wen, J., Chen, L., and Zhao, M., "Optimizing Retail
Systems via Edge Computing," IEEE Transactions on Cloud
Computing, 2023.
[10] Gupta, R., Sharma, M., and Deshmukh, P., "Contactless
Payments in IoT Ecosystems," Journal of Financial
Technology, 2022.
[11] Fouladi, M., Singh, R., and Thomas, L., "Hybrid
Product Detection Systems for Retail," IEEE Transactions
on Retail Automation, 2022. [12] Wilson, J., Parker, S., and
Martinez, C., "Blockchain and IoT Convergence in Retail,"
Journal of Emerging Technologies, 2023.
[13] De Bellis, C., Hildebrand, K., Ito, K., and Herrmann,
A., "Autonomous Shopping Systems: Identifying and
Overcoming Barriers to Consumer Adoption," Journal of
Business Strategy, 2020.
[14] Garcia, L., Fernandez, M., and Lopez, J., "IoT Systems
for Real-Time Billing," International Journal of IoT
Applications, 2023.
[15] Trigo, A., Moreira, J., and Silva, P., "Scalable
Figure 5: Comparison Between Simulation and Hardware Autonomous Retail Solutions," Journal of Smart Commerce,
Implementation 2022.
[16] AiFi Inc., "AiFi Retail in 2024: How Autonomous Store
ACKNOWLEDGMENT Solutions Are Transforming Industries for the Better," AiFi
We sincerely express our gratitude to Prof. Indu K for her Research Publications, 2024.
invaluable guidance, encouragement, and continuous support [17] Szabó-Szentgróti, S., Rámháp, B., and Kézai, P. K.,
throughout this project. Her expertise and insights have been "Systematic Review of Cashierless Stores (Just Walk Out
instrumental in shaping our research and implementation. We Stores): Revolutionizing The Retail," Management and
also extend our heartfelt thanks to Cambridge Institute of Marketing Challenges for the Knowledge Society, vol.
Technology for providing us with the opportunity, resources, 18, no. 2, 2023.
and a conducive learning environment to carry out this work. [18] Huang, L., Zhang, W., and Chen, Y., "Efficient Retail
Checkout with Edge AI," Journal of Artificial Intelligence
REFERENCES
Applications, 2024.
[1] Kumar, D. Mahesh, Boddula Pooja, Dasari [19] Rossi, M., Verdi, L., and Nunes, P., "Customer
Sowmith, Polam Rakesh, and Dugyala Deepthi, 05, Behavior Analysis in Autonomous Stores," Journal of
May 2024: "AI POWERED AUTO BILLING Business Analytics, 2023.
SYSTEM USING RASPBERRY PI."
[20] Shin, Y., Kim, S., and Lee, J., "Consumer perceived
[2] Lee, R. Y., Chai, T. Y., Chua, S. Y., Lai, Y. L., Sim, Y. risk of using autonomous retail technology," Journal of
W., and Haw, S. C., "Cashierless checkout vision system Business Research, vol. 154, pp. 113–126, 2024.
for smart retail using deep learning," Journal of System and
Management Sciences, vol. 12, no. 4, pp. 232-250, 2022.
[3] Palacios, Ricardo, Piguave, Bryan V., Hernandez,
Jefferson, and Abad, Andres G., "Automatic Retail Dataset
Creation with Multiple Sources of Information
Synchronization," in IPTA, October 2023.

[4]Tan, L., Liu, S., Gao, J., Liu, X., Chu, L., and Jiang, H.,
"Enhanced self-checkout system for retail based on improved
YOLOv10," Journal of Imaging, vol. 10, no. 10, 2024.