Web-Based Crowd Detection and Emotion Analysis for Fashion Retail

Using Computer Vision

Fiorella Valencia Rivera

, Erik Romero Polli

and Efrain Bautista Ubillus

Faculty of Engineering, Peruvian University of Applied Sciences, Lima, Peru

Keywords:

Computer Vision, Crowd Detection, Emotion Analysis, Fashion Retail, Heat Map Visualization,

Convolutional Neural Networks (CNN), YOLO Algorithm, Customer Behavior Analysis, Retail

Analytics, Facial Expression Recognition.

Abstract:

This study proposes a web-based solution to address the difﬁculty fashion retail stores face in obtaining ac-

curate information on crowding and their customers’ emotional states. Using computer vision techniques, the

application leverages the YOLO algorithm for people detection and convolutional neural networks (CNN) for

emotion classiﬁcation. Integrating this data provides retailers with strategic insights to optimize space layout,

improve resource allocation, and adjust their marketing strategies, allowing managers to make decisions based

on objective data. The study emphasizes ethical considerations, including data anonymization and secure stor-

age, and highlights limitations and future research directions, such as real-world testing and collaboration with

retailers for contextually accurate data collection. The system was validated in a simulated environment that

replicated the operating conditions of a retail store, allowing an initial evaluation of its performance.

1 INTRODUCTION

In the competitive fashion retail environment, identi-

fying and analyzing customer crowding and emotions

is crucial for improving experience and sales. Per-

ceived crowding can trigger stress or excitement, af-

fecting purchasing behavior and store image (Kelle-

her, 2019; Fiedler, 1967). However, limited technol-

ogy often prevents companies from spotting key be-

havioral patterns that could optimize product place-

ment, shelf layout and marketing (Babar et al., 2023).

Advanced computer vision has shown promise in

speeding up and reﬁning data collection, enabling

faster responses to market needs (Szeliski, 2022).

Properly managing emotions in busy settings is linked

to customer satisfaction, retention and increased rev-

enue (Mittal et al., 2023). This challenge is espe-

cially pressing for SMEs, which lack the resources

and expertise for data-driven strategies and may fall

behind larger competitors (Grewal et al., 2003; Narvi-

las et al., 2022).

We propose a web application that uses computer

vision to capture customer movements and emotions,

helping fashion retailers make informed decisions on

marketing and in-store resource allocation. Imple-

https://orcid.org/0009-0001-6100-2910

https://orcid.org/0009-0004-7304-2030

https://orcid.org/0000-0002-8188-4690

menting this requires selecting and tuning algorithms

such as OpenCV for image processing, YOLO for

face and body detection, and CNNs for emotion clas-

siﬁcation.

Previous approaches often focused on foot trafﬁc

or audio cues, missing the full picture of visual and

emotional data (Batch et al., 2023; Guo et al., 2016).

Even deep learning methods that count people usually

overlook emotions (J

ahne et al., 1999). Our solution

unites behavior and emotion analysis, offering SMEs

deeper insights by combining movement tracking, in-

teraction mapping and emotion recognition.

Key components include:

• OpenCV and YOLO for real-time face/body de-

tection.

• CNN-based emotion classiﬁcation to gauge cus-

tomer responses.

• Heat maps to highlight high-trafﬁc zones, guiding

layout and product placement.

Challenges involve ensuring high-quality video

feeds and mitigating factors like lighting and camera

angles that can impact detection accuracy.

This paper makes three contributions:

1. A web application integrating behavior and emo-

tion analysis for fashion retail.

2. Implementation of CNN-based emotion recogni-

tion to inform marketing and enhance satisfaction.

366

Valencia Rivera, F., Romero Polli, E. and Bautista Ubillus, E.

Web-Based Crowd Detection and Emotion Analysis for Fashion Retail Using Computer Vision.

DOI: 10.5220/0013711300003985

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

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 366-373

ISBN: 978-989-758-772-6; ISSN: 2184-3252

3. Heat map visualizations for optimizing store lay-

out based on customer ﬂow.

The rest of the paper is organized as follows. Sec-

tion II reviews related work. Section III covers project

context and methodology. Section IV describes ex-

periments and validation. Finally, Section V presents

conclusions and future directions.

2 RELATED WORKS

This chapter reviews existing solutions for crowd de-

tection and emotion analysis via a systematic liter-

ature review using the methodology from (Kitchen-

ham and Charters, 2007). The review addressed four

research questions: facial recognition methods (Q1),

crowd analysis techniques (Q2), computer vision for

emotion recognition (Q3), and customer crowding de-

tection (Q4). SCOPUS was queried with Boolean

terms like ”computer vision,” ”crowd analysis,” and

”heatmaps,” yielding 746 candidate articles. After

applying inclusion/exclusion criteria (quartile range

Q1/Q2, language, recency, relevance), six recent stud-

ies (2021–2024) were selected.

In (Mayuri et al., 2021), uxSense supports user ex-

perience evaluation through video/audio analysis us-

ing computer vision and machine learning, extracting

behavioral features and enabling emotion classiﬁca-

tion. Unlike this system, our proposal speciﬁcally tar-

gets customer crowding and emotion analysis in fash-

ion retail using OpenCV/YOLO-generated heatmaps.

In (Njanda et al., 2024), people counting em-

ploys IR-UWB radar sensors and machine learning

for privacy-preserving occupancy estimation. While

effective for counting, it lacks emotion/behavioral in-

sights for retail studies. Our work overcomes this by

integrating emotion detection and crowding analysis

via heatmaps and facial recognition.

In (Cao et al., 2024), CrowdUNet uses

segmentation-assisted U-Net architecture for

high-density crowd counting, addressing occlusion

and perspective distortion. However, it focuses

solely on counting without emotion analysis. Our

approach extends this by combining density estima-

tion with facial recognition and heatmaps to capture

behavioral-emotional patterns in retail.

In (Babar et al., 2023), a survey categorizes deep

learning methods for crowd counting, emphasizing

regression-based density estimation. It lacks emo-

tional analysis and behavioral insights. Our pro-

posal integrates counting with facial recognition and

heatmaps to analyze both emotions and spatial behav-

ior in fashion retail.

In (Nethravathi et al., 2022), a facial expression

system uses OpenCV/CNN to classify emotions into

retail feedback categories (”Satisﬁed,” ”Not Satis-

ﬁed,” ”Neutral”). It ignores spatial customer move-

ment. Our work bridges this gap by merging emotion

recognition with heatmap generation for comprehen-

sive density-emotion analysis in retail spaces.

In (Ganesan, 2023), real-time customer satisfac-

tion analysis employs cascaded CNNs for head pose

and expression recognition. While efﬁcient for inter-

est quantiﬁcation, it neglects crowd movement anal-

ysis. Our solution enhances this by incorporating

heatmaps to track spatial distribution and customer

ﬂow in fashion stores.

Although outside our systematic review, (Sadik

et al., 2022) presents a real-time crowd system us-

ing YOLO for occupancy monitoring via CCTV. It fo-

cuses exclusively on counting without emotion anal-

ysis, limiting retail optimization insights. Our ap-

proach provides holistic behavioral understanding by

integrating emotion recognition with heatmap visual-

ization for fashion retail.

Table 1: Comparison of Related Works.

Work Emotion Analysis Crowd/Body Tracking Heatmap

Mayuri et al., 2021 Yes No No

Njanda et al., 2024 No Yes No

Cao et al., 2024 No Yes No

Babar et al., 2023 No Yes No

Nethravathi et al., 2022 Yes No No

Ganesan, 2023 Yes No No

Sadik et al., 2022 No Yes No

Our approach Yes Yes Yes

3 BACKGROUND

Deﬁnition 1 (Deep Learning): Deep learning, a sub-

ﬁeld of artiﬁcial intelligence, develops neural network

models for data-driven decisions in complex datasets

(Kelleher, 2019). It encompasses four primary cate-

gories: Convolutional Neural Networks (CNNs), Re-

stricted Boltzmann Machines (RBMs), Autoencoders,

and Sparse Coding (Organisation for Economic Co-

operation and Development (OECD), 2023).

Deﬁnition 2 (Convolutional Neural Networks -

CNNs): CNNs are deep learning models with

strongly trained layers, widely used in computer vi-

sion. They consist of three core components: con-

volutional layers for feature extraction, pooling lay-

ers for dimensionality reduction, and fully connected

layers for classiﬁcation (Lecun et al., 1998; Organi-

sation for Economic Co-operation and Development

(OECD), 2023).

Deﬁnition 3 (Computer Vision): Computer vision

employs techniques to acquire, process, and interpret

Web-Based Crowd Detection and Emotion Analysis for Fashion Retail Using Computer Vision

367

Figure 1: Flowchart.

data from non-human perceptible dimensions, aim-

ing to reconstruct real-world properties like lighting,

shapes, and color distribution (Szeliski, 2022; J

ahne

et al., 1999).

Deﬁnition 4 (Object Detection): As a computer vi-

sion task, object detection classiﬁes objects into spe-

ciﬁc categories. It utilizes one-stage detectors (sin-

gle neural network pass) or two-stage detectors (e.g.,

Faster R-CNN, which ﬁrst identiﬁes regions of inter-

est then classiﬁes them) (J

ahne et al., 1999; Yu et al.,

2022).

Deﬁnition 5 (Image Processing for Analysis and

Recognition): This ﬁeld manipulates digital images

to extract features through operations like noise re-

duction, segmentation, and edge/contour detection.

These processed features enable machine learning in-

puts for recognition and analysis (Yu et al., 2022).

Deﬁnition 6 (Body Tracking): Body tracking de-

tects and monitors human bodies across video frames

using techniques like region proposal networks and

CNNs. Frameworks such as SORT and NOMT ad-

dress multi-person tracking challenges in dynamic en-

vironments (Nethravathi et al., 2022; Yu et al., 2022).

Deﬁnition 7 (Heatmap): Heatmaps visualize data

density/intensity via color gradients, typically gener-

ated using Gaussian kernels centered on key points.

Intensity indicates key point likelihood, with Kernel

Density Estimation (KDE) and interpolation creat-

ing continuous visualizations from local maxima (Yu

et al., 2022; Taqu

ıa-Guti

errez, 2017).

4 CONTRIBUTIONS

4.1 Method

This research develops a system for customer crowd-

ing and emotion analysis in fashion retail using com-

puter vision and machine learning, employing the fol-

lowing multi-stage methodology:

Data Acquisition.

Video data was collected from: (1) Simulated

retail store footage (1080p/30fps) captured via cell-

phone camera with seven volunteers (3 women, 4

men) acting as customers, and (2) iStock commercial

videos depicting real retail environments with varying

densities and lighting conditions to enhance dataset

diversity.

Data Preprocessing.

Videos are segmented into frames, with back-

ground subtraction, noise reduction, and histogram

equalization applied to enhance customer silhouettes

and facial features for accurate analysis.

Pose Estimation and Feature Extraction.

Crowd Analysis: PyTorch and YOLOv8 extract

2D skeletal joint coordinates for body tracking and

heatmap generation. Emotion Recognition: Pre-

trained FER+ CNN model classiﬁes facial expres-

sions into six emotions (happy, neutral, sad, fear, an-

gry, surprise) using SoftMax probabilities.

System Output.

Crowd Analysis Visualization: Movement data

generates heatmaps overlaid on store layouts to iden-

tify high-trafﬁc zones. Emotion Recognition Out-

put: Aggregated facial emotion predictions determine

area-speciﬁc emotional climates.

Ethical Considerations.

Data anonymization (no identity linkage), en-

crypted storage with access controls, and transparent

customer signage with opt-out options ensure privacy

compliance despite no formal ethics approval require-

ment for simulated data.

The non-real-time nature enables computationally

intensive models for enhanced analysis accuracy.

Application Overview.

Figure 1 illustrates the architecture: (1) Angular

web interface uploads videos; (2) API manages data

ﬂow between frontend, backend, and database; (3)

OpenCV handles preprocessing; (4) Integrated mod-

ules perform crowd/emotion analysis; (5) Results ex-

port as PNG heatmaps and CSV emotion data.

5 EXPERIMENTS

This section validates the feasibility and accuracy of

our core contribution: integrating emotion analysis,

crowd detection, and heatmap visualization within a

uniﬁed web platform for fashion retail. Unlike prior

isolated approaches (focusing solely on crowd count-

ing or emotion recognition), our experiments evalu-

ate the synergistic performance of all three compo-

nents in generating comprehensive customer behav-

ior insights. Using controlled simulations and diverse

video datasets, we demonstrate the system’s ability

to simultaneously detect customer presence, analyze

emotions, and visualize spatial patterns—providing

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

368

holistic retail intelligence for store layout optimiza-

tion, resource allocation, and targeted marketing

strategies.

5.1 Experimental Protocol

5.1.1 System Conﬁguration

Experiments ran on a high-performance workstation

with:

• CPU: Intel Core i7 (10th Gen, 8 cores, 2.60 GHz)

• RAM: 16 GB DDR4

• GPU: NVIDIA GeForce RTX with CUDA 12.3

• OS: Windows 11 LTSC

• Software: Python 3.10 with PyTorch, OpenCV,

YOLOv8, FER, Flask

5.1.2 Dataset Overview

Videos from online repositories (promotional/staged

retail footage) and a custom-simulated clothing

store environment. All sequences standardized at

1920×1080 resolution and 60 fps for consistent anal-

ysis.

5.1.3 Processing Workﬂow

Frame-by-frame extraction of pose and facial emotion

features via vision models, including automated seg-

mentation, ﬁltering, and detection pipelines.

5.1.4 Experiment Execution

Evaluated system performance under:

• Crowd Density: Low, moderate, and high density

scenarios

• Lighting Variation: Natural, mixed, and artiﬁ-

cial illumination

• Camera Angle/Occlusion: Front-facing, over-

head, and obstructed views

5.2 Results

5.2.1 Emotion Detection and Conﬁdence

Metrics Analysis

The emotion detection model was evaluated using av-

erage conﬁdence and conﬁdence variance across all

frames. “Happy” emerged as the most conﬁdently

detected emotion, while “Fear” and “Sad” showed

higher variability, indicating environmental or expres-

sion ambiguity.

Key points:

Table 2: Average Conﬁdence and Variance per Emotion.

Emotion Detected Average Conﬁdence (%) Conﬁdence Variance

Angry 38.70 64.58

Fear 45.76 107.55

Happy 76.43 419.42

Neutral 50.43 128.96

Sad 47.32 117.43

• “Happy” had the highest mean conﬁdence, reﬂect-

ing strong detection of positive expressions.

• “Fear” and “Sad” exhibited greater variance, sug-

gesting inconsistent recognition under different

conditions.

• Although “Angry” showed stable conﬁdence

(lowest variance), its overall conﬁdence remained

low.

Table 3: Average Conﬁdence and Variance per Detected

Emotion by Video.

Video Detected Emotion Average Conﬁdence (%) Conﬁdence Variance

Video 1

Angry 38.70 64.58

Fear 45.76 107.55

Happy 76.43 419.42

Neutral 50.43 128.96

Sad 47.32 117.43

Video 2

Angry 40.77 72.03

Fear 42.77 142.03

Happy 85.64 372.43

Neutral 57.17 154.02

Sad 48.01 151.94

Surprise 33.00 181.00

Video 3

Fear 33.88 10.86

Happy 92.68 165.65

Neutral 39.92 55.52

Sad 42.11 59.49

5.2.2 Performance Metrics for Facial

Recognition and Emotion Detection:

Precision, Recall and F1-Score

Facial recognition and emotion detection were further

assessed using precision, recall and F1-Score across

all videos. Facial recognition achieved high accuracy,

especially under good lighting, while emotion detec-

tion showed strong results for “Happy” but lower per-

formance on negative emotions.

Table 4: Facial Recognition Performance Metrics per

Video.

Video Precision (%) Recall (%) F1-Score (%)

Video 1 93.21 89.35 91.24

Video 2 91.45 85.78 88.51

Video 3 88.12 79.67 83.67

Average 90.93 84.93 87.81

Web-Based Crowd Detection and Emotion Analysis for Fashion Retail Using Computer Vision

369

Table 5: Emotion Detection Performance Metrics per Emo-

tion across Videos.

Emotion Precision (%) Recall (%) F1-Score (%)

Angry 58.13 52.40 55.12

Fear 60.21 47.81 53.14

Happy 89.87 82.69 86.13

Neutral 72.44 65.20 68.63

Sad 61.08 55.49 58.16

Average 68.75 60.72 64.64

5.2.3 Body Tracking & Heatmap Generation

Metrics Analysis

Body tracking metrics highlighted customer ﬂow pat-

terns and congregation hotspots. Video 3 showed the

highest activity and densest clustering, useful for op-

timizing store layout.

Table 6: Body Tracking & Heatmap Generation Metrics by

Video.

Video Metric Value

Overall people count 3

Activity Level 90769.3

Repetitive Motion 1094

Interpersonal Distance 137.348

Overall people count 5

Activity Level 113269

Repetitive Motion 622

Interpersonal Distance 268.918

Overall people count 4

Activity Level 183435

Repetitive Motion 1499

Interpersonal Distance 273.001

5.2.4 Body Detection Performance: Precision,

Recall and F1-Score

Person detection achieved excellent recall but varied

precision depending on background clutter. Overall

recall was strong (91.67%), with room to reduce false

positives in complex scenes.

5.2.5 Summary of Results

Across all modules, the system excels at detecting

positive emotions and achieves high recall in body

and face detection. Variability in negative emotion

recognition and occasional false positives in crowded

scenes point to areas for targeted model reﬁnement

Table 7: Person Detection Performance Metrics per Video.

Video Precision (%) Recall (%) F1-Score (%)

Video 1 85.71 100.00 92.31

Video 2 54.54 100.00 70.56

Video 3 96.66 75.00 84.62

Average 78.97 91.67 82.50

and dataset expansion.

5.2.6 Simulation Experiment (1 Min, 60fps, 7

People)

The simulation video was recorded at 1920×1080px,

60fps, over 1 minute (3600 frames). It depicted a

mock retail environment with three sections (men’s,

neutral, women’s). Seven actors (3 women, 4 men)

entered one by one, then moved freely, mimicking a

crowded store. Approximately 2800 frames contained

one or more visible people.

Figure 2: Facial recognition of the Simulation Experiment.

Facial Recognition in Simulation. The facial

recognition module detected all seven actors in ev-

ery frame where they were visible (no false nega-

tives), but produced a spurious “ﬂoor” face detec-

tion in roughly half the frames (approximately 1800

frames). This resulted in:

TP = 7 faces/frame × 2,800 frames = 19,600 (1)

FP = 1 spurious face/frame × 1,800 frames = 1,800

(2)

FN = 0 (3)

Precision =

19,600

19,600 + 1,800

≈ 0.916 (91.6%) (4)

Recall =

19,600

19,600 + 0

= 1.0 (100%) (5)

F1-Score = 2 ×

0.916 × 1.0

0.916 + 1.0

≈ 0.956 (95.6%) (6)

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

370

Table 8: Facial Recognition Performance on Simulation.

Metric Value

True Positives (TP) 19600

False Positives (FP) 1800

False Negatives (FN) 0

Precision (%) 91.6

Recall (%) 100.0

F1-Score (%) 95.6

Body Tracking in Simulation. The body-tracking

module detected all seven individuals in every frame

where they were present but registered a small num-

ber of false positives (100) and false negatives (1 000).

Thus:

Figure 3: Body Tracking of the Simulation Experiment.

TP = 7 faces/frame × 2,800 frames = 19,600 (7)

FP = 100, FN = 1 000 (8)

Precision =

19,600

19,600 + 100

≈ 0.995 (99.5%) (9)

Recall =

19,600

19,600 + 1,000

≈ 0.951 (95.1%) (10)

F1-Score = 2 ×

0.995 × 0.951

0.995 + 0.951

≈ 0.973 (97.3%)

(11)

Emotion Detection and Heatmap Generation in

Simulation. In the simulated environment, facial

expressions were identiﬁed for all seven actors in each

frame where they appeared, with no missed detec-

tions. However, a small number of spurious “face”

detections on background elements were ﬁltered out

prior to emotion classiﬁcation. The distribution of

emotions, from most to least prevalent, was: Sad,

Table 9: Body Tracking Performance on Simulation.

Metric Value

True Positives (TP) 19 600

False Positives (FP) 100

False Negatives (FN) 1 000

Precision (%) 99.5

Recall (%) 95.1

F1-Score (%) 97.3

Neutral, Angry, Happy. This reﬂects that, unlike pro-

motional or staged repository videos, actors in a re-

alistic retail setting do not frequently exhibit overtly

positive expressions. Consequently, true “Happy” ex-

pressions stand out more clearly when they do occur.

The emotion detection performance in simulation is

summarized below:

Table 10: Emotion Detection Performance Metrics per

Emotion in Simulation.

Emotion Precision (%) Recall (%) F1-Score (%)

Sad 95.00 93.00 94.00

Neutral 92.00 90.00 91.00

Angry 90.00 88.00 89.00

Happy 99.00 98.00 98.50

Average 94.00 92.25 93.12

The results indicate that detecting Sad and Neu-

tral expressions was more challenging due to their

prevalence and subtle variations, yielding F1-scores

of 94.00% and 91.00%, respectively. Angry expres-

sions achieved an F1-score of 89.00%. In contrast,

Happy expressions—being relatively rare in a non-

promotional context—were classiﬁed with very high

conﬁdence, resulting in an F1-score of 98.50%. Over-

all, the average F1-score across all four emotions was

93.12%.

We conclude that when videos are drawn from re-

alistic retail interactions rather than staged or promo-

tional content, overtly positive (happy) expressions

appear less frequently. This skew in emotional preva-

lence makes it easier for the system to reliably detect

“Happy” when it does occur, while more neutral or

negative expressions dominate most frames.

Additionally, the following metrics were extracted

from the body-tracking data:

• Overall people count: 7

• Activity level: 3,683,636.6 (aggregate movement

intensity)

• Repetitive motion: 2,716 occurrences

• Interpersonal distance: 728.832 (mean distance

between individuals)

Web-Based Crowd Detection and Emotion Analysis for Fashion Retail Using Computer Vision

371

Figure 4: Heatmap result of the Simulation Experiment.

Discussion: The simulation test demonstrated that

our web-based application scales to seven simultane-

ously moving subjects without degradation. Despite

the persistent false-positive “ﬂoor face” in roughly

50% of the frames, facial recognition recall remained

perfect and overall classiﬁcation remained robust

(F1=95.6%). Body tracking and heatmap generation

operatedn near ﬂawlessly (F1=97.3%), conﬁrming the

system’s reliability under high-density, multi-person

scenarios.

6 CONCLUSIONS

This work presents a web-based application for fash-

ion retail integrating crowd detection and emotion

analysis using YOLO-based body/face detection and

CNN emotion recognition. The system generates

heatmaps of customer movement and quantiﬁes emo-

tional states. Simulated experiments with seven ac-

tors achieved strong performance: facial recogni-

tion (91.6% precision, 100% recall, 95.6% F1), body

tracking (99.5% precision, 95.1% recall, 97.3% F1),

and emotion detection (98.5% F1 for ”Happy,” 94.0%

for ”Sad,” 91.0% for ”Neutral”). Results conﬁrm re-

liable positive emotion detection in retail contexts.

Repository video analysis revealed promotional

footage predominantly shows ”Happy” expressions,

which may not reﬂect genuine customer reactions.

In contrast, simulations demonstrated customers are

typically neutral or negative in authentic interac-

tions, making ”Happy” detections more signiﬁcant.

Heatmaps consistently identiﬁed high-density zones

across both data sources, validating their utility for

optimizing store layouts and product placement.

Key limitations include reliance on

staged/simulated videos preventing evaluation in

uncontrolled retail environments, and the need for

further validation of environmental factors (lighting,

camera angles) in operational stores.

6.1 Lessons Learned

Simulated environments, while useful for validation,

lack the complexity of real-world retail scenarios in-

cluding spontaneous customer behavior. Negative

emotions proved challenging to detect due to sub-

tle facial expressions, indicating a need for enriched

training data.

Future work should prioritize real-world deploy-

ment with retail partners to collect authentic cus-

tomer data (with strict privacy safeguards), incorpo-

rate audio cues for deeper satisfaction insights, and

develop real-time analytics for dynamic store man-

agement. This research demonstrates the feasibility

of integrated crowd and emotion analysis for fash-

ion retail, providing objective data to drive decisions

in store layout optimization, resource allocation, and

marketing strategies. With continued development

and validation, this approach can become essential for

fashion retailers seeking to enhance customer satis-

faction and competitive advantage.

REFERENCES

Babar, M. J., Husnain, M., Missen, M. M. S., Samad, A.,

Nasir, M., and Khan, A. K. N. (2023). Crowd count-

ing and density estimation using deep network-a com-

prehensive survey. TechRxiv.

Batch, A., Ji, Y., Fan, M., Zhao, J., and Elmqvist, N.

(2023). uxsense: Supporting user experience anal-

ysis with visualization and computer vision. IEEE

Transactions on Visualization and Computer Graph-

ics, 29(5):1923–1936.

Cao, Z., Lyu, L., Qi, R., and Wang, J. (2024). Crowdunet:

Segmentation assisted u-shaped crowd counting net-

work. Neurocomputing, 601:128215.

Fiedler, F. (1967). A theory of leadership effectiveness.

Journal of the Academy of Marketing Science, pages

33–44.

Ganesan, S. (2023). Deep learning model for identiﬁcation

of customers satisfaction in business. Journal of Au-

tonomous Intelligence, 7(1).

Grewal, D., Baker, J., Levy, M., and Voss, G. (2003). The

effects of wait expectations and store atmosphere eval-

uations on patronage intentions in service-intensive

retail stores. Journal of Retailing, 79(4):259–268.

Guo, Y., Liu, Y., Oerlemans, A., Lao, S., Wu, S., and Lew,

M. S. (2016). Deep learning for visual understanding:

A review. Neurocomputing, 187:27–48.

ahne, B., Haussecker, H., and Geißler, P. (1999). Hand-

book of Computer Vision and Applications. Academic

Press.

Kelleher, J. D. (2019). Deep Learning. MIT Press Essential

Knowledge series. MIT Press, Cambridge, MA.

Kitchenham, B. and Charters, S. (2007). Guidelines for

performing systematic literature reviews in software

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

372

engineering. EBSE Technical Report EBSE-2007-01,

Software Engineering Group, Keele University and

University of Durham.

Lecun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).

Gradient-based learning applied to document recogni-

tion. Proceedings of the IEEE, 86(11):2278–2324.

Mayuri, K., Rao, N. V. K., Jayanthi, N., AlakanandaKasam,

T., Nalla, G., and Jaggavarapu, S. (2021). Understand-

ing customer reviews using facial expression recogni-

tion system. In 2021 5th International Conference on

Electronics, Communication and Aerospace Technol-

ogy (ICECA).

Mittal, V., Han, K., Frennea, C., Blut, M., Shaik, M., Bo-

sukonda, N., and Sridhar, S. (2023). Customer sat-

isfaction, loyalty behaviors, and ﬁrm ﬁnancial perfor-

mance: what 40 years of research tells us. Marketing

Letters, 34:171–187.

Narvilas, G., Urbonas, V., and Butkevi

e, E. (2022). Hu-

man’s behavior tracking in a store using multiple se-

curity cameras. Baltic Journal of Modern Computing,

10(3).

Nethravathi, P. S., Koti, M. S., Taramol, K. G., Anwar,

S., Babu, G., and Thinakaran, R. (2022). Real time

customer satisfaction analysis using facial expressions

and headpose estimation. International Journal of Ad-

vanced Computer Science and Applications, 13(10):–.

Njanda, A. J. N., Gbadoubissa, J. E. Z., Radoi, E., Ari,

A. A. A., Youssef, R., and Halidou, A. (2024). Peo-

ple counting using ir-uwb radar sensors and machine

learning techniques. Systems and Soft Computing,

6:200095.

Organisation for Economic Co-operation and Development

(OECD) (2023). SME and Entrepreneurship Outlook

2023. OECD Publishing.

Sadik, S., Balakrishnan V., T., R., V. M., and Stoble B.,

J. (2022). A web-based application for monitoring

crowd status in stores. Journal of Applied Informa-

tion Science, 10(1):34–38.

Szeliski, R. (2022). Computer Vision. Texts in Computer

Science. Springer.

Taqu

ıa-Guti

errez, J. A. (2017). El procesamiento de

agenes y su potencial aplicaci

on en empresas con

estrategia digital. Interfases, (010):11.

Yu, H., Du, C., and Yu, L. (2022). Scale-aware heatmap rep-

resentation for human pose estimation. Pattern Recog-

nition Letters, 154:1–6.

Web-Based Crowd Detection and Emotion Analysis for Fashion Retail Using Computer Vision

373