Customer behavior analysis and statistics system in intelligent retail scenario

Abstract

The application belongs to the technical field of data processing, and specifically discloses a customer behavior analysis and statistics system in an intelligent retail scene. The system comprises: a multi-dimensional space perception module, which constructs a real-time digital twin space; a social interaction graph construction module, which generates a social correlation graph; a behavior contagion dynamics analysis module, which quantifies behavior diffusion trends; a group behavior semantic analysis module, which identifies consumer preferences and generates a description vector; a retail trend predictability prediction module, which predicts group purchase outbreak probabilities; and a dynamic decision support and feedback module, which automatically generates a closed-loop optimization instruction set. The application quantifies social contagion effects, realizes accurate foresight of retail trends and automatic resource tuning, and improves operational agility.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a customer behavior analysis and statistics system in smart retail scenarios. Background Technology

[0002] Smart retail, a product of the deep integration of artificial intelligence and big data technologies, aims to reconstruct the connections between people, goods, and places in a business environment through digital means, thereby achieving a leapfrog improvement in operational efficiency. The core of this field lies in the real-time perception and intelligent processing of multi-source heterogeneous data within the retail environment. By constructing a refined physical space monitoring system, it provides scientific data support for inventory optimization, precision marketing, and store layout adjustments, thereby enhancing the competitive advantage of physical retail entities.

[0003] Customer behavior analysis and statistics systems in smart retail scenarios are a key branch of the smart retail framework. By integrating computer vision, radio frequency identification (RFID), and multimodal sensing technologies, they achieve automated capture and structured statistics of customer entry, movement patterns, dwell time, and product interaction behavior. The core task of such systems is to analyze customer behavior patterns and potential preferences in specific consumption environments, transforming unstructured physical behaviors into quantifiable business indicators to provide closed-loop feedback for retail decisions.

[0004] While existing technologies have matured in handling individual trajectory tracking and basic behavior recognition, they still present challenges when dealing with dynamic scenarios such as complex promotional activities or new product launches. Traditional analysis systems often focus on feature extraction at the individual level and isolated behavior recording, severely neglecting the social contagion effect among individual customers that drives group purchasing behavior. This results in the system's inability to accurately simulate and predict the diffusion logic of information and emotions within a crowd.

[0005] This lack of social connection dimension makes it difficult for the system to predict potential explosive purchases or evolving consumption trends in the face of sudden surges in customer traffic or word-of-mouth marketing environments. Furthermore, existing statistical models lack sufficient semantic depth in analyzing customer interactions, making it difficult to transform complex social psychological factors into interpretable predictive variables. This results in retailers lacking the ability to respond to fluctuations in group behavior when formulating promotional strategies, leading to inventory imbalances or misallocation of marketing resources. Summary of the Invention

[0006] The purpose of this invention is to provide a customer behavior analysis and statistics system for intelligent retail scenarios, in order to solve the problem that existing technologies focus on extracting individual-dimensional features while ignoring the social contagion effect among customers on the driving role of group purchasing behavior, and to make up for the shortcomings of existing systems in handling complex dynamic scenarios, such as the inability to accurately simulate the logic of information and emotion diffusion and the difficulty in predicting the trend of explosive group purchases.

[0007] The present invention provides a customer behavior analysis and statistics system for intelligent retail scenarios, comprising: The multi-dimensional spatial perception module is used to acquire panoramic visual flow data, radio frequency identification signals and environmental sensor data in the retail physical space in real time, and to construct a real-time digital twin space containing customer location, posture and product location information through multi-source data fusion algorithms. The social interaction graph construction module is used to extract spatiotemporal proximity features, visual joint attention features, and body movement synchronization features between individual customers based on the real-time digital twin space output by the multi-dimensional spatial perception module, so as to generate a dynamic interaction graph that represents the strength of social connections between customers. The Behavioral Contagion Dynamics Analysis Module is used to calculate the diffusion probability of individual behavioral signals in social networks through a behavioral contagion model using a dynamic interaction graph generated by the social interaction graph construction module, in order to quantify the evolutionary trend of specific behaviors from core individuals to the surrounding population. The group behavior semantic parsing module is used to perform deep semantic analysis on the evolutionary trends output by the behavior contagion dynamics analysis module, identify consumption preferences, purchase motives and emotional polarity at the group level, and generate structured group behavior description vectors. The retail trend prediction module is used to predict the probability of group purchase outbreaks and customer flow fluctuation trends within a future preset time period based on the group behavior description vector generated by the group behavior semantic parsing module, combined with historical sales data and current promotional strategies, and using a spatiotemporal evolution neural network. The dynamic decision support and feedback module is used to automatically generate a closed-loop instruction set, including inventory scheduling suggestions, marketing resource allocation plans and store traffic optimization strategies, based on the prediction results output by the retail trend predictive prediction module, and send it to the retail execution terminal in real time.

[0008] Preferably, the multidimensional spatial perception module uses non-biological features to protect privacy during individual identification, constructing a unique temporary feature vector through clothing texture, body shape parameters, and motion features; the temporary feature vector is automatically destroyed after the customer leaves the retail space; the visual processing link of the multidimensional spatial perception module adopts an architecture of edge computing and cloud collaboration, completing target detection and feature extraction at edge nodes, and only uploading structured metadata to the central server.

[0009] Preferably, when determining a collaborative shopping group, the social interaction graph construction module compares the movement trajectory, dwell time, and rhythm of picking up goods between two individual customers using a dynamic time warping algorithm; if the speed coupling degree between the two is higher than 0.85 and the probability of consistency in movement direction is greater than 90%, then they are determined to belong to a collaborative shopping group; the social association strength is a weighted combination of physical proximity, interaction frequency, and synchronization index.

[0010] Preferably, the process by which the behavioral contagion dynamics analysis module calculates the diffusion probability is as follows: Obtain the comprehensive probability that a behavioral signal spreads from the source node to the target node at a specific moment; The calculation of the comprehensive probability is based on the product of the normalized social association weights of neighboring nodes in the social interaction graph, the behavioral intensity coefficient of the core individual, the environmental background correction factor obtained by looking up environmental parameters, the sensitivity of product attributes set based on product categories, and the spatial diffusion damping coefficient determined by real-time Euclidean distance.

[0011] Preferably, the behavioral contagion model adopts a fusion architecture of a linear threshold model and an independent cascade model; The dynamic decision support and feedback module automatically switches the matching model algorithm according to the product category of the current monitoring area. For high-priced, low-frequency decision-making behavior, a linear threshold model is adopted, that is, when the cumulative value of social influence is greater than the individual's specific decision threshold, the contagion is determined to be successful. For low-priced, high-frequency impulsive consumption behavior, an independent cascade model is adopted, that is, each social interaction leads to behavior following according to a preset probability.

[0012] Preferably, the group behavior semantic parsing module is also used to identify group anxiety or expectation, and accordingly send instructions to the retail environment control system to automatically adjust the air conditioning temperature or background music style. The group behavior description vector also includes population profile statistics of segmented consumer groups divided by clustering algorithms. The population profile statistics include age distribution, gender ratio and diversity of group composition.

[0013] Preferably, the preset time period set by the retail trend prediction module includes 15 minutes, 1 hour, or 1 business day; The retail trend prediction module incorporates Monte Carlo simulation, which involves changing the shelf display order, adding on-site experience elements, or adjusting promotional periods within a digital twin space to simulate the evolution logic of group behavior under different strategic interferences.

[0014] Preferably, when generating a marketing resource allocation plan, the dynamic decision support and feedback module uses a multi-objective optimization algorithm to calculate the Pareto optimal solution among improving sales conversion rate, reducing operating costs and optimizing customer experience; The closed-loop instruction set is distributed to each execution unit via a wireless network protocol. The execution unit includes intelligent electronic shelf labels, sales staff handheld terminals, replenishment robots, and automated settlement systems.

[0015] Preferably, the dynamic decision support and feedback module is equipped with emergency avoidance logic; When the system detects fluctuations in group behavior such as running, gathering and conflict, or unusual noise, it suspends all marketing instructions and switches to safety guidance mode, guiding customers to disperse in an orderly manner through in-store broadcasts and signage systems. When the predicted probability of a group purchase surge exceeds the preset warning threshold, the dynamic decision support and feedback module automatically triggers an inventory warning and sends a scheduling signal to the replenishment robot.

[0016] Preferably, it also includes a self-learning evolution unit, which is used to periodically compare the prediction results output by the retail trend predictive prediction module with the actual sales data and customer flow data. The self-learning evolution unit uses a backpropagation mechanism to adjust the infection probability parameters in the behavior contagion dynamics analysis module and the feature weights in the social interaction graph construction module based on the residual values ​​generated by the comparison, so that the system can adapt to changes in customer behavior patterns under different business districts and cultural backgrounds.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a dynamic interaction graph that includes spatiotemporal, visual, and limb synchronization features, deeply exploring the social connections between individuals in retail scenarios and breaking the limitations of traditional systems that only focus on the trajectory of isolated individuals. By introducing the social contagion effect into the behavioral analysis framework, this invention can accurately capture the diffusion logic of information and emotions in a crowd, providing a new scientific perspective for understanding group purchasing decisions.

[0018] 2. This invention utilizes a behavioral contagion dynamics model to quantitatively assess the evolution trend of consumer behavior, significantly improving the prediction accuracy of the system in dynamic scenarios such as complex promotions or new product launches. By calculating the contagion probability and critical point, retail operators can predict the outbreak of group purchases in advance, effectively avoiding inventory imbalances or human resource misallocations caused by information lag, and greatly enhancing the operational agility of retail entities.

[0019] 3. The closed-loop management achieved by this invention, from multi-dimensional perception to dynamic decision feedback, transforms high-dimensional group behavior semantics into executable business instructions. The system can not only predict trends but also optimize resource allocation in real time based on changes in group behavior, including traffic flow adjustment and precise marketing guidance. This intelligent scheduling mechanism based on real-time feedback achieves a dual improvement in retail space utilization efficiency and customer consumption experience, providing a complete systematic solution for the digital transformation of smart retail.

[0020] 4. The self-learning evolution unit and privacy protection mechanism introduced in this invention ensure the robustness and compliance of the system model over a long period of time. By continuously optimizing the infection parameters and feature weights, the system can automatically adapt to the changing business environment. Meanwhile, the desensitization processing and temporary feature vector design balance data value mining and consumer privacy protection, laying a technical foundation for the large-scale deployment and sustainable operation of the system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic interaction graph and behavioral contagion dynamics based on social association in this invention; Figure 3 This is a logical flowchart of the multi-dimensional spatial perception and real-time digital twin space construction in this invention; Figure 4 This is a schematic diagram illustrating the feature evolution of semantic parsing of group behavior and spatial recognition of emotional state in this invention; Figure 5 This is a schematic diagram of the closed-loop feedback control logic for predictive retail trends and dynamic decision support in this invention. Detailed Implementation

[0022] Example 1: Please refer to the appendix Figure 1 This embodiment discloses a customer behavior analysis and statistics system for intelligent retail scenarios. Its overall architecture consists of a complete closed-loop control link, from bottom-level perception and intermediate logic analysis to top-level decision feedback. The system deeply integrates computer vision, IoT communication, dynamics modeling, and deep learning technologies, aiming to achieve accurate prediction and guidance of group consumption behavior through quantitative analysis of complex interactive relationships within the retail space.

[0023] The core operation of the system begins with the multi-dimensional spatial perception module. Please refer to the appendix. Figure 3 This multi-dimensional spatial perception module acquires raw data streams through a sensor matrix deployed at key nodes throughout the retail physical space. The visual sensors include multiple high-resolution cameras mounted on the ceiling and top of shelves, capturing panoramic video streams at 30 frames per second, covering areas including the entrance area, main merchandise display areas, aisles, and checkout areas. An edge computing unit analyzes the video streams in real time, using a deep residual network to extract the skeletal key points of each customer entering the field of view, specifically 18 core coordinate points including the head, shoulders, elbows, wrists, hips, knees, and ankles.

[0024] Simultaneously, a facial feature point extraction algorithm operates to locate the corners of the customer's eyes, mouth, and eyebrows, providing high-dimensional pixel features for subsequent emotion analysis. In addition to visual data, the multi-dimensional spatial perception module also integrates a radio frequency identification (RFID) sensor network. Through an RFID antenna array deployed within the store, the system can capture the signal strength fingerprint emitted by the customer's mobile device. The RFID signal sampling frequency is set to 10 Hz. By analyzing the signal attenuation differences between different antennas, the system uses a trilateration algorithm to initially determine the customer's planar coordinates.

[0025] To eliminate heterogeneity among multi-source data and improve positioning accuracy, the multi-dimensional spatial perception module introduces spatiotemporal alignment and fusion algorithms. In the temporal dimension, the system uses a high-precision network time protocol to ensure that the timestamp error of all sensors is controlled within 5 milliseconds. In the spatial dimension, the algorithm employs weighted least squares to align the visual coordinate system and the RFID coordinate system into a 3D Cartesian coordinate system of the retail space. For occlusion issues occurring in complex stacked areas, the system activates a Kalman filter algorithm, combining historical velocity vectors and acceleration constraints to predictively smooth the individual's motion trajectory.

[0026] When a customer moves from the field of view of one camera to the field of view of another, the cross-camera relay tracking algorithm ensures consistency in individual identification by comparing the individual's clothing texture feature vector and body geometry parameters. The multi-dimensional spatial perception module ultimately outputs a real-time digital twin space. Within this digital twin space, not only are the 3D spatial coordinates and posture sequence of each customer updated in real time, but the precise coordinates, category, and real-time inventory status of all goods within the retail area are also mapped.

[0027] Combined with appendix Figure 2 The generated real-time digital twin space is transmitted to the social interaction graph construction module. The mission of this module is to transform isolated individual data into a graph topology reflecting sociological connections. First, the system calculates the physical distance between any two customers. When this physical distance remains greater than 30 seconds but less than the social safety threshold of 1.5 meters, the system determines that the two individuals have a potential for interaction. Subsequently, the social interaction graph construction module analyzes the shared visual attention characteristics of the two individuals.

[0028] By calculating the orientation vectors of customers' heads and the frequency of eye contact, if both individuals frequently focus their gaze on the same shelf area or the same item, and the cumulative time of eye contact accounts for more than 30% of their time together, the strength of their social connection is significantly increased. Furthermore, this social interaction graph construction module extracts synchronized body movements and uses a dynamic time warping algorithm to compare their movement trajectories, dwell time, and the rhythm of their actions in picking up items. If their speed coupling degree is higher than 0.85 and the probability of consistency in their movement directions is greater than 90%, they are determined to belong to a collaborative shopping group.

[0029] The generated dynamic interaction graph is stored in matrix form. Nodes in the graph represent individual customers, and edges between nodes represent social connections. The edge weights are a weighted composite of physical proximity, interaction frequency, and synchronicity metrics. To capture the instantaneous changes in social relationships within retail scenarios, the social interaction graph construction module employs a sliding window mechanism. The window length is set to 60 seconds, with a step size of 5 seconds. This means the graph is globally updated every 5 seconds, enabling it to keenly identify short-lived social behaviors such as casual conversations or brief following.

[0030] To ensure data integrity, the social interaction graph construction module also includes built-in logic for filtering out false signals. By accessing a pre-defined database of professional behavior characteristics, the system can automatically identify sales staff and cleaning personnel. Sales staff typically exhibit frequent cross-regional movement, fixed directional postures, and short-term interactions with multiple unrelated customers. Once such professional nodes are identified, the system reduces their influence weight in the graph by 70%, ensuring that subsequent analysis focuses on genuine customer social drivers.

[0031] Building upon the social interaction graph, a behavioral contagion dynamics analysis module is introduced. This module aims to quantify how specific behaviors spread like a virus within social networks. Please refer to the appendix for further details. Figure 2 The system first defines the source of infection signals. When a core individual in the dynamic interaction graph exhibits key behaviors, including picking up a product to read the instructions carefully, trying out samples, engaging in lively discussions about product features with peers, or adding a product to their shopping cart, the individual is marked as an infected node. Susceptible customers in the surrounding area who are edge-connected to this infected individual are likely to imitate or deviate from these visual or auditory infection signals.

[0032] The behavioral contagion dynamics analysis module applies an improved probabilistic transformation model to calculate this evolutionary trend. This model considers not only the strength of social associations but also environmental context factors. These factors include the tempo of currently playing background music; a tempo above 120 beats per minute significantly shortens an individual's decision-making hesitation period. Furthermore, warm light color temperatures between 3000 and 4000 Kelvin enhance social interaction activity, and promotional announcements with a sound pressure level between 65 and 75 decibels strengthen information penetration.

[0033] Specifically, the behavioral contagion dynamics analysis module calculates the diffusion probability of behavioral signals using a specific method. The core of this calculation is the comprehensive probability that a behavioral signal will spread from the source node to the target node at a given moment. The calculation process incorporates the normalized social association weights between nodes in the social interaction graph, as well as the relevant information of all adjacent nodes. It also considers the behavioral intensity of the node, which increases sequentially with the depth of actions such as picking up, viewing, trying out, and adding to the shopping cart, with values ​​ranging from 0 to 1.

[0034] In addition, an environmental background correction factor will be considered, which can be obtained by looking up tables based on real-time parameters of music, lighting, and sound pressure levels. Different initial coefficients will be set for high-frequency consumer goods and high-priced electronic products to reflect the differences in the sensitivity of product attributes. Regarding spatial influence, an exponential decay method will be used to simulate the physical characteristic that social influence rapidly weakens with increasing spatial distance. This decay will be calculated by combining the spatial diffusion damping coefficient and the real-time straight-line distance between the two nodes in the digital twin space.

[0035] Through the above calculations, the behavioral contagion dynamics analysis module can generate a behavioral evolution heatmap in real time. The heatmap not only displays the distribution of current purchasing behavior but also indicates potential customer groups that may follow suit within the next 3 minutes through highlighted areas. This forward-looking quantitative assessment provides core data support for identifying group buying sprees.

[0036] Following this, the group behavior semantic analysis module performs a qualitative analysis of the aforementioned evolutionary trends. Please refer to the appendix. Figure 4 This group behavior semantic analysis module transforms low-level coordinate data and probability values ​​into high-level semantic labels. By clustering the micro-expressions of group members using a deep convolutional neural network, the system can identify the overall emotional polarity of the current group.

[0037] For example, when members of a collaborative shopping group frequently exhibit raised eyebrows, upturned lips, and a forward lean, the system labels them as excited; if they show slight furrowing of the brow, frequent flipping of tags, and repeated shifts in body weight, they are labeled as hesitant. This group behavior semantic analysis module constructs a multi-dimensional emotional state space model, combining the convergence of body movements with the co-evolution of facial expressions to generate a structured vector describing group behavior. This vector contains 256 dimensions of features, including group size density, interaction depth indicators, emotional stability coefficients, and attention saturation for specific product categories.

[0038] Subsequently, the retail trend prediction module incorporates the aforementioned high-dimensional features. Please refer to the appendix. Figure 5 This retail trend prediction module employs a long short-term memory network with an integrated attention mechanism. The input layer receives a sequence of vectors describing group behavior, with a time step set to the past 30 minutes. The spatial attention mechanism is responsible for analyzing the flow logic between different shelf areas; for example, it was found that the excitement level in the cosmetics area typically leads to a surge in customer traffic in the perfume area after 15 minutes.

[0039] The time-attention mechanism is responsible for uncovering the periodic patterns of group behavior. By combining historical sales data with patterns from business days and holidays, it predicts the probability of a group purchase surge in the next 15 minutes, 1 hour, or even the entire business day. The prediction process also incorporates Monte Carlo simulations, which rehearse the behavioral diffusion path after changing the intensity of promotions or adjusting shelf layout in a digital twin space, thereby calculating the critical threshold that triggers a group purchase surge.

[0040] Finally, the dynamic decision support and feedback module generates a closed-loop instruction set based on the prediction results. When the predicted probability of a purchase surge exceeds the preset 85% warning threshold, the module automatically sends a replenishment instruction to the inventory scheduling system. Upon receiving the signal, the replenishment robot will plan its route to the designated shelf area according to the optimal path. Simultaneously, the system controls the digital signage in the store, switching the content played from general advertisements to precise promotional content matched with the current high-interest group profile in real time.

[0041] For areas of emotional anxiety detected by the group behavior semantic analysis module, the system will automatically adjust the area's air conditioning temperature to decrease by 1 to 2 degrees Celsius and switch the background music to a soothing style to optimize the consumer psychological atmosphere. In addition, if the system detects abnormal fluctuations in the heat value of customer flow in a local area, such as group running or abnormal gathering and conflict, the emergency avoidance logic will immediately suspend all marketing instructions, switch to safety guidance mode, and guide the orderly dispersal of customers through voice broadcasts.

[0042] Example 2: This example focuses on the application of the system in high-priced, low-frequency decision-making scenarios, such as high-end electronics specialty stores. In this scenario, customer purchasing behavior is often accompanied by a long experience and complex social negotiation, and the logic of behavioral contagion manifests as a linear cumulative effect with a high threshold.

[0043] During the operation of the multi-dimensional spatial perception module, the system significantly improved the accuracy of capturing body details. The camera device focused on the customer's actions in operating the handheld electronic device, including the frequency of screen swiping, the force of pressing physical buttons, and the number of times the device was flipped to observe the base. The RFID sensor network added monitoring of the frequency of the anti-theft cable being stretched on the demonstration unit, serving as a supplementary dimension for multi-source data fusion. In this scenario, the real-time digital twin space not only includes spatial coordinates but also integrates the system interaction logs of the demonstration device.

[0044] For such scenarios, the social interaction graph construction module increases the weight of eye contact features by 40%. Since decisions regarding high-end products often involve in-depth conversations with family members or companions, the system specifically records the duration of eye contact between customers and their partners. If two people frequently engage in verbal interaction and eye contact while viewing the same laptop, their social connection is assessed as extremely strong.

[0045] In the behavioral contagion dynamics analysis module, since high-end products are less likely to trigger impulsive purchases, the system switches to a linear threshold model. This means that a single social signal is insufficient to trigger a purchase intention; contagion is only considered successful when the cumulative value of social influence exceeds an individual's specific decision threshold.

[0046] The specific algorithm logic is as follows: The system initializes a decision resistance value for each customer. When their companions exhibit positive behavioral signals such as taking an instruction manual or consulting a salesperson, these signals, weighted by social factors, are continuously injected into the decision accumulator of susceptible customers. Simultaneously, a decay factor continuously deducts from the accumulated value over time. Only when the accumulated value suddenly surges and breaks through the decision resistance value within a short period will the retail trend prediction module output a high-probability purchase prediction.

[0047] In this embodiment, the group behavior semantic analysis module focuses on identifying hesitation. By analyzing the subtle gait movements of customers pausing in front of products and the rhythm of their fingers tapping on the table, the system can distinguish whether a customer is deeply contemplating or simply waiting. A decision pressure coefficient is added to the group behavior description vector to characterize the degree of consensus within the current group.

[0048] Upon receiving such predictions, the dynamic decision support and feedback module generates a set of instructions that are more inclined towards human intervention and guidance. The system automatically sends push notifications to experienced sales consultants with handheld devices, informing them which group's accumulated decision-making value is approaching the threshold and suggesting the best entry point for intervention, such as providing technical parameter comparisons or explanations of installment payment plans. This precise intervention suggestion can effectively overcome the final obstacle in the purchase decision-making process.

[0049] Furthermore, the self-learning evolutionary unit plays a crucial role in this embodiment. Since the sales sample size for high-end products is relatively small, the unit utilizes a backpropagation mechanism to correct the initial distribution of decision resistance values ​​in the linear threshold model based on actual transaction results or out-of-store data. By comparing the residuals between the predicted purchase breakout point and the actual transaction time, the system continuously optimizes the contagion probability parameters, ensuring that the model can adapt to changes in consumer psychology under different brand premiums.

[0050] Example 3: This example explores the application of the system in fast-moving consumer goods (FMCG) scenarios involving high-frequency, impulse buying, such as promotional displays in large supermarkets. In this case, behavioral contagion exhibits a strong independent cascading characteristic, meaning that even a minor social disturbance can trigger a large-scale follow-up purchase.

[0051] In this scenario, the multi-dimensional spatial perception module focuses on processing high-density passenger flow data. Edge computing architecture excels here, performing the detection and tracking of numerous targets at the front end, uploading only metadata containing individual IDs, coordinates, and basic poses. The visual processing chain utilizes clothing textures to construct unique temporary feature vectors, which are retained in memory for only 4 hours before being automatically destroyed to meet privacy compliance requirements.

[0052] In this embodiment, the social interaction graph construction module significantly relaxes the social safety threshold, limiting physical distance to within 1.2 meters. This is because in a crowded supermarket environment, spatial compression between unrelated customers does not necessarily represent social interaction. The system uses a trajectory consistency algorithm to eliminate nodes that are merely temporary parallel interactions due to spatial congestion. True social association is defined as a group with significantly synchronized picking actions or lingering together in front of a display shelf for more than 15 seconds.

[0053] The behavioral contagion dynamics analysis module switched to an independent cascade model. In this model, each contagion source signal has an independent trigger probability. For example, when a customer takes a box of discounted strawberries from a display stand, as an independent random event, there is a 30% probability that it will directly stimulate susceptible customers within 5 meters to take the same thing.

[0054] The algorithm's principle is described as follows: The system establishes a diffusion chain based on independent event triggers. For each activated behavior node, the system traverses all nodes connected by outgoing edges in the interaction graph. For each adjacent node, the system performs a Bernoulli trial based on a preset infection strength. If the trial is successful, the target node is activated and enters an infected state. This chain reaction can quickly simulate the outbreak of group behavior in scenarios such as limited-time flash sales.

[0055] The group behavior semantic analysis module targets impulsive consumption, focusing on extracting the emotional stability coefficient. In a panic-buying scenario, group emotions typically fluctuate frequently. The system captures the movement rate of group members in front of the shelf; if the movement frequency doubles within two minutes, the area is considered to have entered a high-pressure competition state.

[0056] The retail trend prediction module combines the current rate of inventory reduction with a group behavior description vector, using a spatiotemporal evolutionary neural network to predict the specific time when inventory will be sold out. If the prediction indicates that inventory will be depleted within the next 10 minutes, the dynamic decision support and feedback module will immediately trigger a replenishment robot and simultaneously adjust electronic shelf labels. If replenishment cannot keep up in time, the system will activate a degradation strategy, switching surrounding digital signs to recommendations for relevant alternative products to guide excess purchasing demand to other categories, achieving a dynamic balance in store traffic flow.

[0057] Regarding data security, the multi-source data fusion algorithm does not store any original biometric information during processing. Environmental sensor data, such as conversations captured by acoustic sensor arrays, is only parsed locally for keyword analysis. For example, high-frequency words such as "cheap," "good value," and "fresh" are identified, and the original audio is immediately discarded, retaining only the anonymized semantic tags. This not only reduces the system's storage burden but also protects customers' personal information security from a fundamental technical perspective.

[0058] Example 4: This example focuses on the deep coupling between the system's self-learning evolution unit and privacy protection mechanism. During the long-term operation of the system, the retail environment will change with seasonal transitions, holiday promotions, and significant adjustments to in-store displays. The core task of the self-learning evolution unit is to maintain the real-time effectiveness of each analysis module in the system.

[0059] The self-learning evolutionary unit establishes an optimization loop based on feedback control theory. Please refer to the appendix. Figure 5 The self-learning evolutionary unit periodically retrieves the prediction results from the retail trend predictive module over the past 24 hours and compares them with actual POS machine settlement data and access control passenger flow sensor data in a multi-dimensional manner. A loss function is introduced into the comparison process to quantify the deviation between the predicted trajectory and the actual behavior trajectory.

[0060] For the social interaction graph construction module, if it is found that the predicted social groups exhibit an extremely high split rate in the actual purchase process—that is, predicted to be companions but actually making independent payments with no subsequent connection—the self-learning evolutionary unit will initiate backpropagation, reducing the weight of the social safety threshold in the feature weighting and increasing the weight of the frequency of eye contact. This dynamic adjustment of weights ensures that the graph construction logic can adapt to differences in social distance perception across different regions or cultural backgrounds.

[0061] For the behavioral contagion dynamics analysis module, the unit analyzes whether the predicted behavioral diffusion path matches the actual changes in passenger flow thermal patterns. If the actual spatial diffusion rate of behavior is slower than the predicted value, the unit will reduce the initial coefficient in the contagion probability formula. and increase the spatial diffusion damping coefficient This parameter tuning based on real feedback enables the system to self-evolve across different product categories. For example, during summer cooling mat promotions, the system automatically identifies that the contagion effect of these large items has a longer spatial lag, thereby optimizing its prediction model.

[0062] In terms of privacy protection, the multi-dimensional spatial perception module employs a noise injection mechanism based on differential privacy. Before uploading an individual's movement trajectory to the cloud for large-scale trend prediction, the system adds a small perturbation conforming to a Laplace distribution to the coordinate data. This perturbation does not affect the accuracy of calculating the probability of a group's purchase surge on a macroscopic level, but on a microscopic level, it prevents any third party from reconstructing the precise whereabouts of a specific individual from the cloud data.

[0063] Furthermore, a time-limited key encryption scheme was designed for the feature vectors of clothing textures. The feature vectors extracted by the visual processing link are hashed using a dynamic key that changes every hour. This means that even if the database is illegally accessed at a certain moment, only discrete data fragments that cannot be correlated across time periods will be obtained. When the system determines that a customer has left the retail space for 30 minutes and has not generated any subsequent related behavior, all related features, interaction records, and social edge weights of the individual in the digital twin space are completely erased, and only the statistical mean without individual identification is retained for subsequent report generation.

[0064] This balance between privacy and performance, coupled with the continuous iteration of self-learning evolutionary units, makes the customer behavior analysis and statistics system in smart retail scenarios not only an efficient business auxiliary tool, but also an ethically sound and highly robust intelligent management platform.

[0065] Example 5: This example describes a specific application of the system in optimizing complex customer flow in large retail spaces such as shopping malls. In this scenario, the system needs to handle a larger spatial range, exhibit greater sensor heterogeneity, and coordinate customer flow guidance between different brand stores.

[0066] The multi-dimensional spatial perception module deploys a multi-level visual link within the large shopping mall. The bottom layer consists of wide-angle cameras covering public areas, responsible for capturing large-scale customer flow backgrounds; the middle layer consists of directional cameras targeting specific store entrances, responsible for accurate counting and initial identity screening; and the top layer consists of close-up cameras in the shelf area. To achieve millisecond-level synchronization, the system introduces a fiber optic distributed data interface, ensuring that the latency of visual stream data transmission to the edge server array is less than 2 milliseconds.

[0067] The social interaction graph construction module introduces a hierarchical graph structure within large spaces. The system first constructs a coarse-grained group association graph, identifying large-scale social units such as families, couples, or tour groups. Subsequently, for smaller groups entering specific brand areas, a fine-grained dynamic interaction graph is constructed. By analyzing the consistency of movement of large groups in public spaces, the system can predict their potential behavioral diffusion logic after entering a specific store. For example, families exhibiting high levels of activity and interaction in public areas are far more likely to engage in group purchasing behavior in the children's clothing section than individual customers who are strangers to each other.

[0068] The behavioral contagion dynamics analysis module specifically introduces a cross-regional spillover effect model in this scenario. When a flagship store is conducting a large-scale promotional event, the resulting behavioral signals not only spread within the store's social networks but also radiate to susceptible individuals in the corridors via glass walls or open entrances. The system measures this cross-regional impact through specific quantitative calculations. These calculations consider the total cross-regional influence, the strength of behavioral signals at different locations in the source area, and the specific effect of the signal propagation from the source area to the target area—signal attenuation is affected by the degree of obstruction from physical walls, the width of openings, and the acoustic reflection coefficient of the corridors. The system also clearly defines the area covered by the calculations. Through this quantitative calculation, the system can predict the impact of the flagship store's promotions on the overall floor's customer flow distribution.

[0069] The retail trend prediction module uses a spatiotemporal evolution neural network to analyze bottleneck areas in large-scale circulation routes. By simulating the flow of people at different times and under different weather conditions, the system can identify which areas are prone to excessive congestion that leads to social disruption.

[0070] The dynamic decision support and feedback module generates real-time adjustment strategies for the floor wayfinding system based on predicted traffic flow trends. If a surge in crowd activity is predicted on the south side of the second floor, while the north side is relatively quiet, the system automatically adjusts the digital signage in public areas, prioritizing displays of brand offers for the north side to dynamically smooth out peak and off-peak traffic. In emergencies, such as fire alarms or the discovery of suspicious packages, the dynamic decision support and feedback module can immediately take over all commercial display devices. Based on real-time digital twin spatial data, it calculates the shortest and least congested escape route for each gathered group and provides directional guidance via a voice array.

[0071] This embodiment fully demonstrates the system's collaborative processing capabilities in ultra-large-scale, multi-variable environments, proving that by capturing social contagion effects, it can not only increase sales but also achieve fundamental optimization of physical space operational efficiency.

[0072] Example 6: This example details the emergency and extreme performance of the system during specific holiday promotions, such as the handling strategy when customer traffic surges to more than 5 times the normal level during Double 11 offline experience days or anniversary celebrations.

[0073] Under these extreme conditions, the multi-dimensional spatial perception module automatically activates a load balancing mechanism. Edge computing nodes dynamically redistribute computing tasks according to the complexity of the area. For the central square area with extremely high pedestrian traffic, the system switches to a lightweight object detection model, which, although sacrificing some keypoint accuracy, ensures real-time tracking without frame drops. The visual link adopts adaptive hierarchical rendering technology, prioritizing the extraction speed of coordinates and basic poses, while postponing the task of subtle emotion analysis to an asynchronous processing queue during off-peak hours.

[0074] The social interaction graph construction module incorporates a dense subgraph recognition algorithm. In extremely crowded scenarios, where physical distance becomes ineffective, the system extracts the consistency of individual action frequencies to isolate social relationships. For example, in a crowd of shoppers, only nodes exhibiting synchronized movement, synchronized arm outstretching, and continuous dialogue are clustered as social units. The dense subgraph recognition algorithm effectively prevents the overgeneralization of social connections due to excessive crowd density, ensuring that the edge weights of the graph accurately reflect the driving forces of consumption.

[0075] The behavioral contagion dynamics analysis module employs a rapid propagation simulation based on particle swarm optimization during peak periods. By abstracting customer behavior as particles with inertia and social attraction, the system can simulate tens of thousands of possible diffusion paths in a very short time. The prediction time window is shortened to 5 minutes to cope with rapidly changing market feedback. The system focuses on monitoring super-node customers with high spillover effects. Once such customers post their on-site experience on social media or send strong recommendations to their companions via voice, the system immediately increases their behavioral weight within the area, predicting potential localized surges in customer traffic.

[0076] In this embodiment, the group behavior semantic analysis module focuses on stress detection. By analyzing the anxiety index in the facial expressions of group members and the frequency of body swaying, the system can identify whether the group is in a state of irrational panic buying.

[0077] In this scenario, the primary task of the dynamic decision support and feedback module is to ensure safety and order. When the system predicts that the crowd pressure coefficient in a certain shelf area has reached a critical value, it will automatically trigger a flow control command. The smart fences in the store will adjust their physical angle according to the command to slow down the flow of people entering the area. At the same time, the system sends a path avoidance signal to the replenishment robot to ensure that the robot does not cause congestion in the crowd, but instead completes the remote delivery of goods through a preset hidden maintenance channel.

[0078] After peak periods end, the self-learning evolutionary unit performs offline deep learning on the massive amounts of data generated. By comparing prediction biases under extreme loads, the system can learn the nonlinear abrupt changes in consumer psychology under extreme pressure. For example, it discovers that during periods of extreme crowding, customers become less sensitive to price but more sensitive to speed of acquisition. This knowledge is incorporated into the system's core algorithm library, enabling the system to have stronger predictive capabilities for similar events in the following year.

[0079] This embodiment demonstrates that the system described in this invention not only possesses sophisticated analytical capabilities but also exhibits strong engineering robustness and adaptability to extreme scenarios, enabling it to meet the demands of high-intensity retail real-world operations. Through multi-level feedback control and flexible model switching, the system successfully transforms complex sociological behavioral patterns into quantifiable, predictable, and controllable retail management data, greatly promoting the in-depth development of intelligent retail towards scientific and digital transformation.

[0080] In summary, this invention constructs an intelligent retail behavior statistics system capable of understanding and utilizing social contagion effects through deep coupling of multi-dimensional perception, graph construction, dynamic analysis, semantic parsing, predictive models, and decision feedback. The system not only overcomes the limitations of traditional systems that rely on isolated analysis of individual characteristics, but also improves the operational efficiency and security of retail spaces through closed-loop management, demonstrating both commercial application value and technological foresight.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A customer behavior analysis and statistics system for intelligent retail scenarios, characterized in that, include: The multi-dimensional spatial perception module is used to acquire panoramic visual flow data, radio frequency identification signals and environmental sensor data in the retail physical space in real time, and to construct a real-time digital twin space containing customer location, posture and product location information through multi-source data fusion algorithms. The social interaction graph construction module is used to extract spatiotemporal proximity features, visual joint attention features, and body movement synchronization features between individual customers based on the real-time digital twin space output by the multi-dimensional spatial perception module, so as to generate a dynamic interaction graph that represents the strength of social connections between customers. The Behavioral Contagion Dynamics Analysis Module is used to calculate the diffusion probability of individual behavioral signals in social networks through a behavioral contagion model using a dynamic interaction graph generated by the social interaction graph construction module, in order to quantify the evolutionary trend of specific behaviors from core individuals to the surrounding population. The group behavior semantic parsing module is used to perform deep semantic analysis on the evolutionary trends output by the behavior contagion dynamics analysis module, identify consumption preferences, purchase motives and emotional polarity at the group level, and generate structured group behavior description vectors. The retail trend prediction module is used to predict the probability of group purchase outbreaks and customer flow fluctuation trends within a future preset time period based on the group behavior description vector generated by the group behavior semantic parsing module, combined with historical sales data and current promotional strategies, and using a spatiotemporal evolution neural network. The dynamic decision support and feedback module is used to automatically generate a closed-loop instruction set, including inventory scheduling suggestions, marketing resource allocation plans and store traffic optimization strategies, based on the prediction results output by the retail trend predictive prediction module, and send it to the retail execution terminal in real time.

2. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 1, characterized in that, The multidimensional spatial perception module uses non-biological features to protect privacy during individual identification, constructing a unique temporary feature vector through clothing texture, body shape parameters, and motion features; the temporary feature vector is automatically destroyed after the customer leaves the retail space; the visual processing link of the multidimensional spatial perception module adopts an architecture of edge computing and cloud collaboration, completing target detection and feature extraction at edge nodes, and only uploading structured metadata to the central server.

3. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 2, characterized in that, When determining a collaborative shopping group, the social interaction graph construction module compares the movement trajectory, dwell time, and rhythm of picking up goods between two individual customers using a dynamic time warping algorithm. If the speed coupling degree between the two is higher than 0.85 and the probability of consistency in movement direction is greater than 90%, then they are determined to belong to a collaborative shopping group. The social association strength is a weighted combination of physical proximity, interaction frequency, and synchronization index.

4. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 3, characterized in that, The process by which the behavioral contagion dynamics analysis module calculates the diffusion probability is as follows: Obtain the comprehensive probability that a behavioral signal spreads from the source node to the target node at a specific moment; The calculation of the comprehensive probability is based on the product of the normalized social association weights of neighboring nodes in the social interaction graph, the behavioral intensity coefficient of the core individual, the environmental background correction factor obtained by looking up environmental parameters, the sensitivity of product attributes set based on product categories, and the spatial diffusion damping coefficient determined by real-time Euclidean distance.

5. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 4, characterized in that, The behavioral contagion model adopts a fusion architecture of linear threshold model and independent cascade model; The dynamic decision support and feedback module automatically switches the matching model algorithm according to the product category of the current monitoring area. For high-priced, low-frequency decision-making behavior, a linear threshold model is adopted, that is, when the cumulative value of social influence is greater than the individual's specific decision threshold, the contagion is determined to be successful. For low-priced, high-frequency impulsive consumption behavior, an independent cascade model is adopted, that is, each social interaction leads to behavior following according to a preset probability.

6. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 5, characterized in that, The group behavior semantic parsing module is also used to identify group anxiety or expectation, and accordingly send instructions to the retail environment control system to automatically adjust the air conditioning temperature or background music style. The group behavior description vector also includes population profile statistics of segmented consumer groups divided by clustering algorithms. The population profile statistics include age distribution, gender ratio and diversity of group composition.

7. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 6, characterized in that, The preset time period set by the retail trend prediction module includes 15 minutes, 1 hour, or 1 business day. The retail trend prediction module incorporates Monte Carlo simulation, which involves changing the shelf display order, adding on-site experience elements, or adjusting promotional periods within a digital twin space to simulate the evolution logic of group behavior under different strategic interferences.

8. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 1, characterized in that, When generating a marketing resource allocation plan, the dynamic decision support and feedback module uses a multi-objective optimization algorithm to find the Pareto optimal solution among improving sales conversion rate, reducing operating costs and optimizing customer experience. The closed-loop instruction set is distributed to each execution unit via a wireless network protocol. The execution unit includes intelligent electronic shelf labels, sales staff handheld terminals, replenishment robots, and automated settlement systems.

9. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 1, characterized in that, The dynamic decision support and feedback module is equipped with emergency avoidance logic; When the system detects fluctuations in group behavior such as running, gathering and conflict, or unusual noise, it suspends all marketing instructions and switches to safety guidance mode, guiding customers to disperse in an orderly manner through in-store broadcasts and signage systems. When the predicted probability of a group purchase surge exceeds the preset warning threshold, the dynamic decision support and feedback module automatically triggers an inventory warning and sends a scheduling signal to the replenishment robot.

10. The customer behavior analysis and statistics system in the intelligent retail scenario according to claim 1, characterized in that, It also includes a self-learning evolution unit, which is used to periodically compare the prediction results output by the retail trend predictive prediction module with the actual sales data and customer flow data. The self-learning evolution unit uses a backpropagation mechanism to adjust the infection probability parameters in the behavior contagion dynamics analysis module and the feature weights in the social interaction graph construction module based on the residual values ​​generated by the comparison, so that the system can adapt to changes in customer behavior patterns under different business districts and cultural backgrounds.

Images