In-Store Customer Behavior Analytics Using Computer Vision

Introduction

I

n the dynamic landscape of the contemporary economy, the retail sector has undergone profound structural changes driven by growing competition, persistent changes in consumer preferences, diversification of shopping channels, and rising expectations regarding shopping experience quality (Kostyal et al., 2026). Under these complex conditions, firms’ survival and sustainable competitive advantage are increasingly linked to their ability to understand customer behavior deeply, accurately, and in real time (Garrido-Moreno et al., 2024)). Key retail marketing decisions—including store layout optimization, shelf placement, comprehensive customer experience (CX) management, in-store promotional campaigns, and loyalty strategies—depend on a sound understanding of how consumers encounter and interact with the store’s physical environment (Lemmens et al., 2026). However, the empirical and theoretical literature suggests that traditional consumer research methods cannot fully meet these emerging demands. Historically, market research has relied heavily on self-report methods and conventional qualitative instruments, including in-depth interviews, focus groups, and structured questionnaires. Although relatively convenient and inexpensive, these instruments have substantial methodological limitations when representing actual consumer behavior (Salih et al., 2025; Theocharis et al., 2025). Previous studies indicate that consumers are often unable to clearly articulate their feelings, perceptions, and mental processes concerning products (Nazari Ghazvini et al., 2026). Moreover, implicit memory and unconscious perceptions are inaccessible to conventional measurement techniques, while interviewer bias and retrospective recall can distort verbal reports. The discrepancy between what customers say and what they actually do at the point of sale creates a substantial intention–behavior gap, which may contribute to misguided marketing decisions and product failures (Ghazvini et al., 2024).

This issue is particularly important when examining consumer decision-making within physical environments. A considerable share of purchases, particularly impulse purchases, involves immediate, intuitive, and sometimes unconscious decisions affected by the store’s visual and spatial stimuli (Rodrigues et al., 2021). A store is therefore not merely a venue for exchanging goods; it is an ecosystem in which experiences are produced, perceptions are shaped, and visual attention may be converted into purchase intentions. Nonetheless, comparisons between physical and online retail reveal a pronounced information asymmetry to the disadvantage of physical retailers (Cameron et al., 2025). Digital platforms can monitor clickstream data, page dwell time, and navigation routes, enabling real-time personalization and fine-grained experience optimization. Physical retailers, by contrast, often depend on retrospective outputs such as point-of-sale (POS) receipts and enterprise resource planning (ERP) systems (Li et al., 2023). Consequently, managers have limited visibility into what happens while customers walk through aisles, stop at shelves, touch or inspect merchandise, or encounter the store’s cold and hot zones.

Overcoming this information gap and the limitations of conventional methods requires objective, continuous, AI-based approaches (Nazari et al., 2023). Computer vision and image processing offer new possibilities for unobtrusively recording actual customer behavior (Castilla et al., 2024). Existing closed-circuit television cameras or intelligent vision sensors can be used to track movement trajectories, quantify dwell time in different zones, estimate visual engagement with particular shelf tiers, and identify physical actions such as picking up merchandise (Knof et al., 2024). Integrating image-derived behavior with customer relationship management (CRM) and transactional data may convert raw surveillance observations into strategic consumer knowledge. Despite expanding interest in digital transformation and smart retail, much of the literature concerns online settings, and relatively few studies comprehensively examine how image-based behavioral analytics relate to key marketing constructs such as satisfaction, engagement, and loyalty within physical stores (e.g., (Yang et al., 2025)). Operational challenges—including illumination changes, camera viewpoints, occlusion, and crowding—also require carefully formulated analytical frameworks if dependable evidence is to support layout redesign, spatial productivity, and conversion-rate improvements. Accordingly, the central research question is: How can intelligent image processing objectively capture and analyze customers’ physical behavior within stores, and how can the resulting patterns inform optimized marketing decisions and novel loyalty models? This study aims to establish an image-processing analytical framework for measuring in-store behavioral variables, including movement paths, stopping locations, and shelf engagement, to support evidence-based retail management.

The remainder of this article is organized as follows. Section 2 reviews the theoretical background and prior research on in-store behavior and computer vision. Section 3 describes the research methodology and proposed system architecture. Section 4 presents implementation findings and analyses. Section 5 discusses managerial implications and conclusions.

Literature Review and Theoretical Background

Over recent years, customer behavior analysis in shopping and commercial environments has moved beyond traditional, one-dimensional approaches toward the integration of artificial intelligence, cognitive indicators, and spatial analytics. Iranian research suggests that intelligent tools are crucial for uncovering complex and nonlinear consumption patterns. In a systematic review, Safinia and Ghavami (2024) found that combining the Internet of Things as an infrastructure for collecting environmental big data with AI as an interpretive tool provides a promising way to identify emotional and impulsive purchasing patterns (Safinia & Ghavami, 2024). Similarly, Gerami et al. (2023), examining deep learning in online stores, addressed data cleaning and the separation of purposeful activities from anomalies, reporting that careful behavioral-data preprocessing significantly increases the accuracy of models predicting purchase tendencies (Gerami et al., 2022). Kazemi and Sadatian (2025) developed a multilayer recommendation framework integrating digital behavioral profiling, machine learning, and behavioral economics (Kazemi & Sadatian, 2025). They showed how combining behavioral traces with intelligent algorithms could reduce cognitive decision errors and improve service personalization, reinforcing the need for structured behavioral data management also highlighted by Kouhzadi et al. (2022) (Kouhzadi et al., 2022).

Understanding visual perception and unconscious reactions forms another strand of the literature. In an eye-tracking experiment, Kimasi and Khoshnevis (2021) found that matching visual stimuli to individual characteristics improves objective attention indicators (Kimasi & Khoshnevis, 2021), such as total attention and gaze duration, although the relationship is moderated by need for cognition. Zahedi Noghabi et al. (2022), drawing on the eye–mind hypothesis, likewise found that individual characteristics and learning styles affect fixation duration and concentration, with deeper cognitive processing associated with longer and more stable fixations (Zahedi Noghabi et al., 2022). Neuromarketing investigations extend these explanations to brain activity. Homayounfar et al. (2023), using electroencephalography (EEG), reported a positive association between visually induced emotional arousal and message recall (Homayounfar et al., 2022). Shirkhodaei et al. (2023), examining the P300 component, found that visual stimuli play a more prominent role than auditory information in mental processing (Shirkhodaei et al., 2023). These findings underscore the relevance of visual information for understanding consumer attention and choice.

Visual attention also manifests in the physical store environment, where architecture and spatial organization shape purchasing behavior. Andalib et al. (2023) used grounded theory to model environmental sensory stimuli and concluded that visual and physical cues, interacting with spatial design, influence behavioral outcomes, including purchasing and loyalty (Andalib et al., 2022). At a broader spatial scale, Yaghoubi et al. (2021) used spatial interaction models and geographic information systems (GIS) to describe customer behavior as a function of geographic location and environmental attractiveness (Yaghoubi et al., 2021). Earlier data-mining studies, including Beshardoust and Asgharizadeh (2023), combined K-means clustering with Apriori association rules to optimize product placement according to complementary purchases; however, those analyses relied solely on post-purchase transaction records (Beshardoust & Asgharizadeh, 2022).

Despite these theoretical and methodological developments, three significant gaps remain in the domestic literature. First, most studies using AI and machine learning focus exclusively on websites and online retailers, overlooking behavioral dynamics in physical stores. Second, studies of visual attention and cognitive behavior are generally conducted in controlled laboratories using wearable instruments such as eye trackers or EEG headsets, which may have limited external validity in dynamic retail environments. Third, field studies either rely on self-report and qualitative techniques or analyze only POS data, leaving shopping paths, stopping locations, and visual shelf interactions largely unmeasured. This study therefore applies image processing to develop an objective, real-time, non-intrusive framework for capturing, quantifying, and analyzing behavior in the physical retail environment.

Alongside domestic research, the international literature offers a more developed account of the convergence of marketing with new vision technologies. Whereas much of the domestic work emphasizes online channels or instrument-based experiments, recent international studies increasingly deploy automated, real-time, unobtrusive systems within physical stores. The field has progressed rapidly from early geometric and statistical models toward deep learning, RGB-depth vision, edge computing, and the Internet of Things. An important starting point was recognition that conventional POS databases do not register pre-purchase preferences. Liu et al. (2016) devised a surveillance system classifying six shelf-facing customer behaviors, ranging from passing without interest to touching, inspecting, and placing goods in a basket (Liu et al., 2017). By integrating head orientation, body orientation across eight directions, and hand actions using dynamic Bayesian networks, they distinguished casual passage from meaningful product attention. To capture such behavior in real time, Liciotti et al. (2017) developed an integrated arrangement of vision and depth sensors with processing embedded locally on cameras, contributing to edge computing and connecting spatiotemporal shelf interactions to managerial decision-support systems (Liciotti et al., 2016).

The assessment of affective and psychological states subsequently became another dimension of shopping-experience analytics. Generosi et al. (2018), noting the limitations of questionnaires and self-report, proposed a system using biometric observations and deep-learning facial-expression analysis to monitor emotional reactions at retail touchpoints (Generosi et al., 2018). Alfian et al. (2019) combined NoSQL databases with big-data frameworks for real-time behavioral processing and used association rules to examine links between store dwell time and purchase conversion (Alfian et al., 2019). Paolanti et al. (2020) introduced the VRAI framework, using overhead cameras and parallel convolutional networks to improve accuracy and offer standardized datasets for person counting, re-identification, and shelf-interaction recognition (Paolanti et al., 2020). To address the sensitivity of end-to-end models to specific store geometries and lighting, Wen et al. (2021) proposed a modular hierarchical system that recognizes activities at different levels of abstraction and adapts to altered store configurations without retraining the entire network (Wen et al., 2021).

More recently, combining computer vision with IoT and cloud-computing infrastructures has expanded operational possibilities. Narvilas et al. (2022) evaluated cloud AI platforms and pretrained models, demonstrating the potential for accurate customer tracking and density heat maps with relatively limited training (Narvilas et al., 2022). Alipio et al. (2023) combined ultra-wideband (UWB) localization with single-shot detector (SSD) algorithms in a real supermarket to identify frequent routes and count human traffic in real time, reporting accuracy above 99% (Alipio et al., 2023). Zangana (2025) combined YOLOv8 for human detection and Mask R-CNN for item-boundary segmentation, overlaying shelf-planogram compliance with customer occupancy heat maps (Zangana, 2025). Yang et al. (2025) combined YOLOv8, DeepSORT, and computational geometry to automatically extract shoppers’ movement vectors and stopping points within large retail stores, improving continuity in crowded scenes (Yang et al., 2025). From a marketing perspective, Rastogi et al. (2025) grouped these technologies into facial affect assessment, dynamic personalization, and automated shelf inventory management (Rastogi et al., 2024). Nguyen et al. (2024) integrated mathematical behavioral models and deep neural networks to visualize individual and group micro-behavior and optimize movement routes (Nguyen et al., 2024). Ganokratanaa et al. (2024) combined YOLOv8 and DeepFace to extract estimated age, gender, and density indicators for reconsidering underused store areas (Ganokratanaa et al., 2024).

In parallel, researchers have investigated fundamental computational and geometric constraints. Gapski (2024) modeled geometric relationships between cameras and applied vision transformers to constrain spatial-localization and body-pose estimation errors under occlusion. Elhoseny et al. (2024) addressed computational demands on low-power devices using lightweight architectures, stacked autoencoders, and metaheuristic optimization, reporting 98.56% accuracy for complex human-behavior recognition on edge hardware without expensive centralized servers (Elhoseny et al., 2025). Memon et al. (2026) surveyed surveillance image processing, from contrast correction, edge detection, and optical flow under unstable lighting to deep learning for anomaly detection (Memon et al., 2026). Finally, Laskar et al. (2025) and AlZahrani (2026) reviewed opportunities such as immediate shopping personalization and automated detection of stock shortages, alongside barriers related to data privacy regulation, large labeled datasets, and the limited interpretability of deep neural networks (AlZahrani, 2026; Laskar et al., 2025). They identified edge computing, synthetic data, and multimodal AI as important directions for next-generation smart retail.

Methods and Materials

In terms of purpose, this is an applied and developmental study. Data collection and analysis follow an objective, non-intrusive observational approach relying on computer vision and spatiotemporal data. Its epistemological foundation is a contemporary positivist, data-driven behavioral paradigm: rather than relying on traditional self-report, retrospective recall, or questionnaires subject to perceptual bias, actual shopper actions—movement vectors, walking speed, stopping points, shelf dwell time, and spatial exploration—are measured as objective, quantitative variables. To enhance ecological validity, the investigation was carried out in an operating retail store. Its physical layout was divided, according to merchandising and planogram principles, into target zones: the entrance threshold and initial decision area, high-traffic principal aisles, special shelves and promotional displays, and checkout and payment areas.

Continuous video streams were captured using existing ceiling-mounted surveillance cameras directed obliquely downward, with Full HD resolution and a conventional video frame rate, providing adequate coverage without disrupting normal shopping activity. The automated processing pipeline comprised four connected steps. First, adaptive frame preprocessing stabilized illumination contrast and reduced reflections from glass shelving. A geometric homography then transformed perspective-distorted image pixel coordinates into real two-dimensional metric coordinates on the store floor plan, calibrating distances, displacements, and velocities in meters. Second, a single-stage deep convolutional neural network detected shoppers in individual frames. Multi-target tracking algorithms combining Kalman filtering with deep appearance descriptors maintained temporal continuity and assigned each individual a unique identifier, reducing identity switches caused by occlusion, crowds, and intersecting walking paths. Third, the resulting trajectories were used to derive complete entrance-to-exit paths, stopping points and shelf proximity dwell time, instantaneous speed, and spatial occupancy density; these features were stored as structured time series.

To convert technical observations into actionable management information, computed indicators were mapped to theoretical constructs from consumer behavior. The proportion of passersby stopping at a shelf was defined as the shelf attractiveness rate; time spent within the active shelf zone represented behavioral engagement and attention; mean motion and acceleration vectors represented shopping pace; trajectory continuity indicated traffic flow and spatial discoverability; and low-occupancy regions on spatial heat maps were classified as physical blind spots. For system validation, automated person counts and dwell time were compared with simultaneous manual observation by independent raters, and standard multi-object tracking and identity matching measures were assessed. To address research ethics and shopper privacy, distinguishing faces and biometric features were blurred and anonymized locally during preprocessing. Processing was restricted to the geometric behavioral descriptors, without retaining raw video recordings.

Findings and Results

The computer-vision system generated behavioral data through continuous monitoring of shoppers, spatiotemporal matrix extraction, and transformation of observed trajectories into objective marketing indicators. During the one-week field observation period, covering peak and off-peak hours, 3,480 unique customer tracking identifiers were recorded. After exclusion of incomplete tracks, defined as presence for fewer than 10 seconds at the entrance threshold, 3,210 valid, complete movement trajectories were available for analysis.

Stage 1: Technical performance and data validity. Before examining marketing indicators, detection and multi-object tracking performance was compared with manually annotated observations over randomly selected time windows comprising 600 seconds of high-traffic footage.

1. Shopper detection accuracy: Detection performance was assessed by mean average precision at an intersection-over-union threshold of 50% (mAP@0.50); the result was mAP@0.50 = 91.4%.

2. Multi-object tracking accuracy (MOTA): The following equation was used:

MOTA = 1 − [Σₜ(FNₜ + FPₜ + IDSWₜ)] / [Σₜ GTₜ]

Here FN denotes missed detections, FP false positives, IDSW identity switches during visual occlusion, and GT the ground-truth count. MOTA was 86.8%. The identity consistency score (IDF1) was 88.2%, suggesting reliable tracking under crowding.

Stage 2: Movement dynamics and spatial discoverability

(a) Walking speed and shopping pace. The instantaneous velocity of shopper i was calculated from calibrated ground-plane positions (x, y) and the frame interval Δt using the following expression:

vᵢ(t) = √[(xᵢ,t+Δt − xᵢ,t)² + (yᵢ,t+Δt − yᵢ,t)²] / Δt

Entrance threshold zone: Mean walking speed was 1.18 ± 0.21 m/s.

Main merchandise aisles: On approaching shelves, mean walking speed declined markedly (p < 0.001); reaching 0.58 ± 0.16 m/s. This reduction was interpreted as a shift from locomotor navigation toward cognitive processing and visual product evaluation.

(b) Physical blind-spot analysis. The walkable store floor was partitioned into 0.5 × 0.5 m cells. Occupancy density (D_c) for cell c was calculated as follows:

D꜀ = [Σᵢ₌₁ᴺ Tᵢ,꜀] / T_total

In this equation, the relevant quantity is the time customer i spent in cell c. The analysis indicated that 21.8% of total store area had an occupancy density below 5% (D_c < 0.05), suggesting poor discoverability of distant side aisles and inefficient utilization of floor space.

Stage 3: Shelf attractiveness and attention-capture rate

(a) Initial attractiveness rate (AR). For each shelf or zone k, the number of shoppers who stopped nearby (N_stop) was divided by the number who passed within the shelf’s visible area (N_pass), as follows:

ARₖ = (N_stop,k / N_pass,k) × 100

A stop was defined as a decline in speed to v ≤ 0.2 m/s for at least t ≥ 2 s.

(b) Behavioral engagement depth (DT). Continuous mean dwell time was measured within the active contact zone of each shelf (less than 0.6 m). The results for different store zones are shown below.

To test whether eye-level and lower shelves differed in dwell time, an independent-samples t-test was conducted.

As Table 2 shows, shopper dwell time at eye level was significantly higher than at lower shelf levels. Cohen’s d indicated a very large effect; on average, shoppers spent approximately 3.6 times longer at eye-level shelves than at bottom shelves. Several parenthetical numerical expressions in the original narrative are left blank, but the corresponding values are reported in Table 2.

Stage 4: Customer behavior in checkout queues and checkout friction. Time spent in the checkout zone was observed as an indicator of friction and potential dissatisfaction. The results for 2,840 customers who reached payment are reported below.

According to Table 3, total mean checkout waiting and payment time was 126.40 seconds (SD = 43.80). Pearson correlations showed a significant positive association between queue delay and friction-related actions, such as shifting in place, motor restlessness, temporarily leaving, or abandoning the queue (r = 0.54, p < 0.01), particularly beyond 90 seconds. Waiting time was also positively associated with visual engagement with impulse-purchase shelves adjacent to checkout stations (r = 0.42, p < 0.01). The source states that each additional 30 seconds of waiting could increase attention to and purchase of impulse items by up to 18%, provided that total waiting time remained below the critical threshold of 180 seconds.

Stage 5: Associations between computer-vision indicators and behavioral loyalty. Finally, movement-track data for loyalty-card holders (N = 640) were matched to their purchase frequency. Multiple linear regression was used to predict repeat purchasing from environmental behavioral variables.

Overall model fit indicators:

R = 0.678; R² = 0.460; adjusted R² = 0.457; F(3, 636) = 180.50; p < 0.001.

Standardized coefficients indicated that store spatial coverage ratio was the strongest positive predictor of behavioral loyalty (β = 0.41, p < 0.001), implying that customers who explored more store zones tended to purchase more frequently during subsequent periods. Total dwell time also had a positive significant association with repeat purchases (β = 0.22, p < 0.001). In contrast, checkout delay was negatively associated with repeat purchases (β = −0.28, p = 0.002), highlighting the adverse role of checkout friction in long-term loyalty.

Discussion

This study examined the feasibility of objectively analyzing shopper behavior in an operational retail store. Computer vision and image processing were used to extract motion patterns, stopping behavior, shelf interactions, and checkout queue behavior. The results showed that image-based observations can go beyond customer counts and movement maps to produce measures associated with attention, behavioral engagement, spatial discoverability, and behavioral loyalty. Image processing can therefore connect observed in-store behavior to managerial and marketing decisions.

Movement analysis revealed reduced customer walking speed upon entering principal aisles and encountering shelves. This change may indicate a transition from spatial navigation to environmental inspection and visual information processing. This interpretation is consistent with Zahedi Noghabi et al. (2022), who related cognitive processes to gaze and attention patterns (Zahedi Noghabi et al., 2022). Homayounfar et al. (2023) and Shirkhodaei et al. (2023) likewise emphasized visual stimuli in emotional activation and mental processing (Homayounfar et al., 2022). The present study extends these issues from laboratory environments to an operational store and suggests that changes in movement can function as behavioral signs of exposure to environmental cues. The result is also consistent with Yang et al. (2025) and Nguyen et al. (2024), who used movement paths and stopping points to study in-store customer behavior (Nguyen et al., 2024; Yang et al., 2025).

The shelf results indicated an association between product placement and height and the frequency and duration of shopper stops. Eye-level shelves and aisle end-caps attracted more engagement, whereas lower tiers and standard longitudinal shelving showed less behavioral engagement. This is in line with Kimasi and Khoshnevis (2021) on visual properties and attention and with Liu et al. (2016) and Paolanti et al. (2020) on shelf interaction (Kimasi & Khoshnevis, 2021; Liu et al., 2017; Paolanti et al., 2020). Those studies show that stopping and product engagement can help distinguish ordinary movement through space from meaningful product interest. Nevertheless, stopping and dwell time alone do not demonstrate visual attention or purchase intention; they should be interpreted as approximate behavioral proxies rather than direct measurements of cognition.

Spatial heat maps also showed that customer presence was unevenly distributed and that some parts of the floor were rarely visited. Identifying these regions can inform layout revisions and improved aisle discoverability. This observation agrees with Andalib et al. (2023), concerning spatial design and environmental stimuli, and Yaghoubi et al. (2021), concerning spatial interactions. It is also related to Zangana (2025) and Ganokratanaa et al. (2024), who discussed heat maps for identifying low-traffic locations and improving store organization (Andalib et al., 2022; Ganokratanaa et al., 2024; Yaghoubi et al., 2021; Zangana, 2025). Occupancy mapping thus helps managers recognize which parts of a store receive limited exposure or traffic; nevertheless, explanations for low traffic should also account for layout, merchandise assortment, signs, and route characteristics.

The checkout findings presented a dual picture of waiting time. Longer waits were associated with greater friction and restlessness, but queue time could also increase exposure to merchandise displayed near payment stations. This is consistent with discussions of impulse purchasing and environmental influences, including Safinia and Ghavami (2024) (Safinia & Ghavami, 2024). However, the result should not be interpreted as support for longer queues: exposure to nearby merchandise does not necessarily produce purchases and can coincide with deterioration in customer experience. Queue management should therefore balance exposure to complementary items against the prevention of friction and dissatisfaction.

In the behavioral loyalty model, store spatial coverage had the strongest positive association with repeat purchases. Total dwell time was also positively associated with repurchasing, whereas checkout delay had a negative association. These patterns suggest that broader exploration of the store is associated with more frequent return visits and that payment-stage friction may relate inversely to repeat purchasing. The findings correspond to Alfian et al. (2019), who used behavioral data to examine conversion, and to Kouhzadi et al. (2022) and Kazemi and Sadatian (2025), who emphasized combining behavioral traces with analytical and CRM systems (Alfian et al., 2019; Kazemi & Sadatian, 2025; Kouhzadi et al., 2022). Compared with transaction-only studies such as Beshardoust and Asgharizadeh (2023), the findings highlight the value of pre-purchase and in-store navigation data. However, the regression establishes statistical associations rather than causal effects. Shopping missions, product assortment, customer characteristics, and overall store experience may also contribute to these relationships.

Methodologically, this study illustrates the application of computer vision to measuring customer behavior under natural retail conditions. Such methods can help address the limitations of self-reports and exclusively transactional records by providing information on movement trajectories, stops, and spatial distribution. This approach aligns with international work on in-store tracking and behavior analysis, including Liciotti et al. (2017), Paolanti et al. (2020), and Narvilas et al. (2022) (Liciotti et al., 2016; Narvilas et al., 2022; Paolanti et al., 2020). Nevertheless, interpretation depends on reliable detection and tracking, camera calibration, illumination, occlusion, and clearly defined behavioral metrics. Privacy protection, limits on retained data, and transparent communication about video-based monitoring must also be core requirements.

Overall, trajectories, stopping points, shelf interactions, and checkout behavior can provide a more detailed account of the shopping experience and inform evidence-based decisions about merchandising, space management, and payment processes. The key advantage lies in observing actual behavior in natural conditions and comparing patterns across store zones. Image-based indicators should nevertheless be interpreted alongside sales records, loyalty data, and supplementary evaluations to avoid overstating cognitive or causal interpretations of observed actions.

Recommendations

Store managers are advised to use heat maps and dwell-time indicators to identify low-traffic sections, reconsider store layout, and improve route signage. In view of the observed differences across shelf positions, product placement should be reviewed in relation to commercial objectives, product category, and desired visibility, rather than placing every product at the same shelf height or position. Checkout waiting times and customers’ behavioral signals should be monitored continuously so that exposure to complementary merchandise is maintained without increasing restlessness or undermining shopping experience.

Future studies should combine image-derived information with transactional records and CRM systems to investigate more accurately the relationship between store exploration, actual purchases, and customer return behavior. Studies across stores with varying sizes, layouts, and product categories could assess generalizability. Subsequent research should validate image-derived indicators using human observation or complementary instruments and examine the roles of shopping mission, crowding, and environmental characteristics. In all such applications, shopper privacy, minimization of identifiable data, and transparency in collection and processing should remain fundamental principles.

Authors’ Contributions

All authors equally contributed to this study.

Declaration

In order to correct and improve the academic writing of our paper, we have used the language model ChatGPT.

Transparency Statement

Data are available for research purposes upon reasonable request to the corresponding author.

Acknowledgments

We would like to express our gratitude to all individuals helped us to do the project.

Declaration of Interest

The authors report no conflict of interest.

Funding

According to the authors, this article has no financial support.

Ethics Considerations

The study protocol adhered to the principles outlined in the Helsinki Declaration, which provides guidelines for ethical research involving human participants.