Intelligent store LED display cabinet screen content pushing method based on image recognition technology

Abstract

The invention relates to the technical field of image recognition, in particular to a store LED display cabinet screen content intelligent pushing method based on the image recognition technology, and the method comprises the steps: collecting a store environment image in real time through a wide-angle camera, and recognizing a face orientation angle and a body orientation angle of a customer; calculating a real-time attention weight in combination with the spatial linear distance between the customer and the LED screen; fusing and aggregating the attention and the commodity thermodynamic diagram data, and generating content pushing priority scores of all the screens; and based on the scoring result, matching the multimedia content, and generating and sending a content updating instruction data packet to drive the target LED screen to perform content refreshing. According to the invention, accurate matching and content intelligent scheduling among multiple customers, multiple screens and multiple commodities can be realized, and the correlation and response efficiency of store display content are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a method for intelligently pushing screen content of store LED display cabinets based on image recognition technology. Background Art

[0002] Amid the rapid development of new retail and smart marketing, store display screens have gradually transitioned from static posters to dynamic LED display cabinets. Deploying LED screens in display areas to play product promotional videos, interactive content, or real-time advertisements can enhance customer immersion, increase product awareness, and boost sales conversion rates. However, faced with diverse display combinations involving multiple categories, multiple products, and multiple screens, as well as the high frequency of changes in customer behavior, traditional static carousels or timed switching strategies struggle to meet the demands for personalization, diversification, and immediate responsiveness.

[0003] Existing technologies suffer from the following major technical bottlenecks: First, most display systems rely solely on fixed content carousels or product priority lists, lacking the ability to perceive on-site customer attention behaviors in real time, making it difficult to tailor content to individual customers. Second, while some systems incorporate headcount or dwell time detection, they often overlook more distinctive behavioral characteristics such as customer facial orientation, body posture, and spatial distance. This results in crude attention judgments and large errors in push results. Third, in the content scheduling and screen-driven processes, there is still a lack of a complete closed-loop processing mechanism for content missing, resource adaptation, and priority conflicts, resulting in insufficient system stability and flexibility. Summary of the Invention

[0004] The present invention provides a method for intelligently pushing screen content on store LED display cabinets based on image recognition technology. This method takes into account perception accuracy, scoring logic, and drive integrity to improve display effects and operational efficiency.

[0005] The method for intelligently pushing content on store LED display cabinet screens based on image recognition technology includes the following steps: S1: A wide-angle camera deployed on the roof of the store collects environmental image data in real time, identifies the facial and body orientation angles of all customers in the image, and outputs customer status data; S2: Calculate the customer's real-time attention weight for each LED screen based on the customer status data and the linear distance between the customer and each LED screen, where the attention weight is negatively correlated with the distance and decays exponentially; S3: Integrate the attention weights of all current customers and superimpose the preset product heat map data to generate the content push priority score for each LED screen; S4: According to the content push priority score, multimedia content corresponding to the product is matched from a preset content library, and a content update instruction is sent to the target LED screen with the highest priority score.

[0006] Optionally, the S1 includes: S11: Real-time environmental image data is collected through a wide-angle camera deployed on the top of the store; S12: Input the environmental image data into the YOLOv7 model to detect the human body bounding boxes of all customers in the image; S13: Based on the human body bounding box, extract the coordinates of the skeleton key points of each customer through the HRNet model; S14: Calculate the angle between the hip joint key point connection vector and the normal direction of the target LED screen according to the hip joint key point connection vector in the skeleton key point coordinates, and output the body orientation angle.

[0007] Optionally, the S1 further includes: S15: Detecting the coordinates of the customer's facial feature points within the human body bounding box area using the RetinaFace model; S16: Calculate the angle between the vector connecting the pupil center and the nose tip in the facial feature point coordinates and the normal direction of the target LED screen, and output the facial orientation angle; S17: Aggregate the body orientation angles and facial orientation angles of all customers to generate customer status data including each customer ID and its corresponding angle value.

[0008] Optionally, the S2 includes: S21: extracting the facial orientation angles and body orientation angles of all current customers from the customer status data; S22: The linear distance between customers and each LED screen is obtained in real time through a group of UWB positioning base stations pre-installed on the store floor; S23: Calculate the cosine value of the angle between the customer's facial orientation vector and the normal direction of the target LED screen according to the facial orientation angle to generate a facial orientation factor; S24: Calculate the cosine value of the angle between the customer's body orientation vector and the normal direction of the target LED screen according to the body orientation angle to generate a body orientation factor; S25: Calculating a distance attenuation factor based on the spatial straight-line distance by applying an exponential function; S26: Based on the facial orientation factor, the body orientation factor and the distance attenuation factor, the customer's real-time attention weight to the target LED screen is calculated.

[0009] Optionally, the attention weight is calculated as: Attention weight = (face orientation factor + 0.5 × body orientation factor) × distance decay factor.

[0010] Optionally, the S3 includes: S31: For the same LED screen, the output attention weights of all customers are accumulated and summed to generate the aggregated attention of the screen; S32: Retrieve pre-stored product heat map data from the store backend system in real time. The product heat map data includes the weight score of each product corresponding to the screen; S33: Multiplying the aggregated attention by the weight score of the corresponding product in the product heat map data to obtain the basic content push priority score of each LED screen; S34: When the difference between the basic content push priority scores of two or more LED screens is less than a preset threshold, the score weight of the screen associated with the high inventory warning level product is increased to generate a final content push priority score.

[0011] Optionally, the weight score of the screen corresponding to the product is generated by weighted calculation of historical customer stay time, touch interaction frequency and inventory warning level.

[0012] Optionally, the S4 includes: S41: Filtering the LED screen corresponding to the highest score from the content push priority scores to generate a target LED screen identifier; S42: According to the product ID associated with the target LED screen identifier, searching for multimedia content bound to the product ID in a preset content library, and outputting matching multimedia content; S43: dynamically adapting and rendering the matching multimedia content to the physical size parameters of the target LED screen to generate a content update instruction data packet that can be parsed by the screen; S44: Sending a content update instruction data packet to the hardware device corresponding to the target LED screen identifier via the store local area network to trigger a refresh of the screen display content.

[0013] Optionally, in S42, when there is no matching content in the preset content library, the default promotional video of the same category of products is automatically called, the backup interactive game program library is activated, and a content missing alarm is sent to the background system.

[0014] Beneficial effects of the present invention: This system, through a multi-stage collaborative processing mechanism employing a wide-angle camera, YOLOv7, HRNet, and the RetinaFace model, can simultaneously achieve high-precision, real-time recognition of the body and facial orientation angles of multiple customers within a store. Compared to traditional solutions that rely solely on position or infrared sensing, this system accurately determines a customer's gaze direction and interaction intent without disturbing them, providing a more reliable basis for behavioral judgment for subsequent push notifications and effectively reducing misplaced content display.

[0015] This invention introduces a three-factor weighted model: facial orientation, body orientation, and spatial distance attenuation. Combined with the weighted scores of corresponding products in the store's backend product heat map data, it establishes a real-time dynamic matching mechanism across multiple customers, multiple screens, and multiple products. This mechanism not only quantifies customer attention across screens but also adjusts priorities based on factors like inventory and interactive behavior. This effectively couples displayed content with customer behavior, improving content exposure efficiency and marketing conversion rates.

[0016] This invention builds an automated, fault-tolerant content-driven system, encompassing a complete control chain from "target LED screen identifier" positioning to "multimedia content matching," "content rendering instruction generation," and "screen driver execution." In particular, in the event of missing content, the system automatically calls promotional videos for the same category, activates interactive game programs, and reports missing content alerts, ensuring uninterrupted display on terminals. This significantly improves the continuity of the user experience and the intelligent level of system operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention; Figure 2 Schematic diagram of the S2 process of an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0020] It should be noted that references in the specification to "one embodiment," "an embodiment," "exemplary embodiments," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not necessarily every embodiment will include such specific features, structures, or characteristics. Furthermore, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0021] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0022] like Figure 1-Figure 2 As shown, the method for intelligently pushing content on store LED display cabinet screens based on image recognition technology includes the following steps: S1: A wide-angle camera deployed on the roof of the store collects environmental image data in real time, identifies the facial and body orientation angles of all customers in the image, and outputs customer status data, specifically: S11, Environmental Image Data Collection: Several wide-angle cameras are evenly distributed throughout the store's ceiling. These cameras use CMOS image sensors with a resolution of at least 1920 × 1080 and a frame rate of 30 fps. The cameras are set to face down vertically or at an angle of no more than 45° to view the entire display case area, ensuring coverage of all customer activity areas in front of the LED screens. The cameras upload real-time image data to edge computing devices or local servers via wired Ethernet or wireless transmission modules.

[0023] S12, human bounding box detection: The collected environmental image data is input into the pre-loaded YOLOv7 model for multi-target detection.

[0024] YOLOv7 is a single-stage object detection neural network based on an anchor box mechanism. Its backbone network utilizes the Efficient Layer Aggregation Networks (ELAN) architecture, resulting in high accuracy and real-time performance. The model is pre-trained on the COCO dataset and fine-tuned to the store environment. It outputs bounding boxes for all customers in each frame. The bounding box data includes the upper-left corner coordinates (x, y), width, and height.

[0025] S13, human body key point extraction: Based on the human body bounding box detected in S12, the HRNet (High-Resolution Network) model is input into each customer image area to obtain the coordinates of the skeleton key points.

[0026] HRNet maintains high-resolution feature maps throughout the network, enabling parallel extraction of features at multiple scales and cross-scale connections through a multi-resolution fusion module. The model outputs the coordinates of 17 key points of each patient's standard skeleton (such as the top of the head, neck, shoulders, elbows, hips, knees, ankles, etc.), returning the two-dimensional coordinates and confidence values ​​for each key point.

[0027] S14, body orientation angle calculation: First, based on the skeletal keypoint coordinates output by the HRNet model, the coordinates of each customer's left and right hip keypoints are extracted and recorded as: Left hip key point coordinates: ; Coordinates of the right hip key point: ; Construct the customer's body orientation vector based on these two key points , defined as the two-dimensional vector pointing from the left hip to the right hip: ; Then, according to the installation direction of the LED screen in the image coordinate system, its normal direction vector is preset as .

[0028] For example, for a screen perpendicular to the bottom of the image, the normal direction can be set to .

[0029] Next, calculate the angle between the body orientation vector and the screen normal direction vector, which is defined as the customer's body orientation angle This angle is calculated using the standard cosine angle formula: ; in, is the dot product of two vectors, is the Euclidean norm of the body orientation vector, calculated as: ; The norm of the screen normal direction vector (usually a unit vector with a value of 1).

[0030] The final output For a value range in A real number. A smaller angle indicates that the customer is facing the screen more directly, while a larger angle indicates a more significant deviation. A threshold angle (e.g., 30°) can be set to determine "valid orientation."

[0031] S15, facial feature point detection: In order to further identify the customer's facial orientation, the facial image area is extracted within the human body bounding box area detected in S13, usually limited to the upper half of the bounding box, especially the head and neck area.

[0032] The image of this region is fed into the RetinaFace model for facial landmark detection. RetinaFace is a high-precision face detection and keypoint localization model based on RetinaNet. Its backbone network uses ResNet-50 or MobileNet, and it utilizes a Feature Pyramid Network (FPN) to enhance multi-scale face detection. The model is pre-trained on the WideRFACE dataset and exhibits strong robustness.

[0033] RetinaFace output includes the two-dimensional pixel coordinates of five standard facial feature points, namely: Left eye center , right eye center, nose tip , left corner of the mouth, right corner of the mouth.

[0034] In the present invention, the left eye center point and the nose tip point are selected as key references for determining the customer's facial orientation.

[0035] S16, facial orientation angle calculation: Based on the feature point coordinates extracted in S15, construct the customer's facial orientation vector. Define vector is a two-dimensional vector pointing from the center of the left eye to the tip of the nose, expressed as: ; Then, using the same LED screen normal direction vector as S14 , calculate the angle between the face facing vector and the screen direction vector.

[0036] Next, we can use the cosine angle formula to get the customer's facial orientation angle. , calculated as: ; represents the dot product of the face orientation vector and the screen direction vector, is the modulus of the face orientation vector, representing the length of the straight line from the pupil to the tip of the nose, is the modulus of the normal vector of the LED screen.

[0037] The final output For a value range in A real number. The smaller the value, the more the customer's face is facing the screen; the larger the value, the greater the degree of deviation.

[0038] S17, customer status data generation: the body orientation angle of each customer Angle with face Perform structured integration to generate customer status data.

[0039] Each customer is assigned a unique identification ID in the system, for example, the customer ID is "C001".

[0040] The system updates the status data of all current customers in a cycle of 1 second, and maintains the angle history records of the last few frames (for example, the last 10 seconds) in the cache to provide support for the subsequent dynamic smoothing of attention weights.

[0041] S2: Based on the customer status data and the linear distance between the customer and each LED screen, calculate the customer's real-time attention weight for each LED screen. The attention weight is negatively correlated with the distance and decays exponentially. Specifically: S21, customer status data analysis: extract structured fields from customer status data. Each customer status data record contains customer ID, facial orientation angle, and body orientation angle .

[0042] S22, spatial straight-line distance acquisition: Based on a group of UWB positioning base stations deployed on the store floor, walls, and around the display area, the spatial coordinates of each customer relative to each LED screen are obtained with an accuracy better than 30cm.

[0043] The customer's three-dimensional space coordinates are ; The three-dimensional coordinates of the LED screen are marked as .

[0044] According to the Euclidean distance formula between two points in three-dimensional space, calculate the Customer and The straight-line distance between the LED screens : ; This distance serves as input to the subsequent distance decay factor calculation.

[0045] S23, face orientation factor calculation: based on the face orientation angle obtained in S21 , that is, Customers facing The angle of the LED screen is calculated to calculate the corresponding face orientation factor .

[0046] The facial orientation factor is defined as the absolute value of the cosine of the angle, which is used to reflect the degree of gaze. The calculation formula is as follows: ; in The unit is degree (°), which must be converted to radians before cosine calculation.

[0047] when When , it means the face is completely facing the screen, the factor is close to 1, and when When , it indicates side view, and the factor is close to 0.

[0048] S24, body orientation factor calculation: Similarly, according to the body orientation angle obtained in S21 , that is, The customer's body is facing The angle of the LED screen is used to calculate the corresponding body orientation factor , defined as follows: ; This value is used to determine whether the customer's body is facing the target LED screen. As the temperature approaches 0°C, the factor approaches 1; when the body is completely deviated (such as facing away), the value approaches 0.

[0049] S25, distance attenuation factor calculation is then based on the straight-line distance between the customer and the LED screen. , introduce the attenuation function to construct the distance attenuation factor .

[0050] In the present invention, the distance decay factor is expressed using an exponential decay model: ; in, is the straight-line distance between the customer and the LED screen (meters), is the distance attenuation coefficient, which is a positive real number and is obtained from the preset "spatial attenuation coefficient comparison table" according to the store type, such as: Convenience store type: ; Supermarket type: ; Exhibition type: ; This factor is used to reflect the objective rule that customers have stronger perception and higher attention to screens that are closer.

[0051] S26, attention weight synthesis: The above three factors are weighted and synthesized to calculate the Customers to The attention weight of LED screens , the synthesis formula is as follows: ; Among them, the weight of the face orientation factor is set to 1, and the weight of the body orientation factor is set to 0.5, indicating that their influence is relatively minor. The distance attenuation factor serves as an overall multiplier to control the overall strength.

[0052] Attention weight The value range is between [0,1], the larger the value, the Customers to The more attention a LED screen receives, the more attention it will receive. Combined Calculate and cache data for subsequent S3 priority score calculation.

[0053] S3: Integrate the attention weights of all current customers and superimpose the preset product heat map data to generate the content push priority score for each LED screen, specifically: S31, customer group attention aggregation: obtain the attention weights of all customers to each LED screen at the current moment from step S2, recorded as ,in Index for customers, Number the LED screen.

[0054] For each LED screen , sum up the attention weights of all customers to generate the aggregate attention of the screen , the calculation formula is as follows: ; in, For the The aggregated attention of LED screens, is the total number of customers identified in the current frame, For the Customers to The attention weight of each LED screen. The higher the value, the higher the collective attention of the current customer group to the screen.

[0055] S32, Product Heat Map Data Acquisition: Retrieve product heat map data associated with each LED screen from the store backend system. This product heat map data is a scoring system built based on historical customer behavior and product operation data, recording the weighted score of each product on its respective display screen. This weighted score is generated by weighting the following factors: Historical customer stay duration: Based on the average gaze time detected by the camera, it is used to reflect visual attention; Touch interaction frequency: refers to the frequency with which customers perform interactive operations such as clicking and sliding on the screen, used to assess the degree of active interest; Inventory warning level: reflects the current inventory status of the product. It is divided into high, medium and low levels according to the degree of inventory tension, and is assigned different scoring values ​​such as 1.0, 1.7 and 0.4 respectively.

[0056] Multiple factors are synthesized through a weighted formula to generate a weighted score between each pair of products and screens. , calculated as follows: ; in, Indicates the length of stay. Indicates the frequency of interaction, Indicates the inventory warning level, is the preset weighting coefficient, satisfying , which can be customized according to store operation strategy.

[0057] This score is used to characterize the "degree to which each product content should be pushed" and, together with the user attention score, determines the final priority.

[0058] S33, basic priority score calculation: aggregate attention and weighted scoring Perform product operation to obtain the basic content push priority score of each LED screen ; in, For the Basic content push priority score for each LED screen, This score is assigned to the weight associated with the screen. This score reflects the combined value of customer attention behavior and product operation needs.

[0059] S34, multi-screen conflict priority correction: To avoid the situation where multiple LED screens have similar scores but different actual marketing priorities, a multi-screen conflict priority correction mechanism is designed.

[0060] First, determine whether the basic score difference between any two screens among all current LED screens is less than the set threshold (like ; ; When the above conditions are met, it means the screen With screen There is a conflict. At this point, the inventory warning level of the product associated with each screen will be further retrieved For screens associated with products with a higher inventory warning level (e.g. level = 1.0), a stock priority correction factor is multiplied. (like ), generate the final content push priority score ; ; For the other screens, leave the rating unchanged: ; Ultimately, all LED screens The sorted data are used as a basis for executing the content push strategy in step S4.

[0061] S4: Based on the content push priority score, the multimedia content of the corresponding product is matched from the preset content library, and a content update instruction is sent to the target LED screen with the highest priority score. Specifically: S41, target screen determination: obtain the content push priority scores of all current LED screens from step S3, All priority scores are compared, and the LED screen with the highest score is selected as the target screen for content push. A corresponding target LED screen identifier is generated to uniquely identify the physical display terminal. This identifier is an internal ID code, such as "LED_05," which is bound to the screen's control address, installation location, and product information.

[0062] S42, Product Multimedia Content Matching: The system searches for product IDs associated with the target LED screen identifier and performs a content matching search within the system's pre-set content library. This pre-set content library is a structured media database that stores multimedia resources such as product videos, carousels, brand introductions, and interactive animations, indexed by product ID.

[0063] If the product ID exists in the content library, all multimedia content bound to it will be extracted and output as a matching multimedia content set. The types include but are not limited to: Promotional video in high-definition MP4 format; PNG / JPEG carousel sequence; HTML5 interactive product introduction component.

[0064] If the product ID does not match any content in the content library, the following fault tolerance strategy will be automatically activated: Call the default promotional video for products in the same category: The system searches for backup videos in the default resource library under the product category tag (such as "beverages," "cosmetics," and "electronics"). Activate the backup interactive game program library: If no promotional video is available, a casual game program will be randomly selected from the interactive entertainment content library, such as brand mini-games and lucky draw interactions, to attract customers to stop; Send missing content alerts to the backend system: Generate missing content alert data packets containing product IDs, timestamps, and screen identifiers, and upload them to the backend content management system to prompt operations and maintenance personnel to supplement the missing content in a timely manner.

[0065] S43, content rendering instruction generation: matching the multimedia content with the physical parameters of the target LED screen, including: Resolution (e.g., 1920×1080); Display ratio (such as 16:9, 1:1, etc.); Available display area (avoiding the part blocked by the border); Refresh rate and frame rate compatibility.

[0066] The system uses a multimedia rendering engine (such as FFmpeg or a proprietary adapter) to process raw footage in real time, including operations like video cropping, image compression, and layer shuffling, to generate a content update instruction packet that complies with the target LED screen's decoding specifications. This packet includes the encoded video stream or image sequence, rendering frame rate and resolution identifiers, a screen receiving end identification field, and content playback control instructions (play / pause / loop, etc.).

[0067] The data packet format can be a binary stream, RTMP push stream frame, or structured control commands that comply with manufacturer protocols.

[0068] S44, screen driver execution: Sends a content update instruction packet via the store's local area network (LAN) to the physical hardware controller corresponding to the target LED screen identifier. Data packet transmission uses the TCP / IP protocol to ensure low latency and high stability.

[0069] After receiving the content update instruction data packet, the screen controller automatically completes the following operations: Decode media content and cache it locally; Reset the original playback status and load new content; Start content refresh and play or display in carousel immediately; Content looping and status switching are performed according to preset playback control strategies (e.g., rotation after 5 minutes of playback, or switching to static images if no customers stay).

[0070] At this point, the screen content update process is completed, and the system will enter the next cycle of priority calculation and content drive waiting state, forming a continuous closed loop.

[0071] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0072] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An intelligent push method for store LED display cabinet screen content based on image recognition technology, characterized in that: The following steps are involved: S1: A wide-angle camera deployed on the roof of the store collects environmental image data in real time, identifies the facial and body orientation angles of all customers in the image, and outputs customer status data; S2: Calculate the customer's real-time attention weight for each LED screen based on the customer status data and the linear distance between the customer and each LED screen, where the attention weight is negatively correlated with the distance and decays exponentially; S3: Integrate the attention weights of all current customers and superimpose the preset product heat map data to generate the content push priority score for each LED screen; S4: According to the content push priority score, multimedia content corresponding to the product is matched from a preset content library, and a content update instruction is sent to the target LED screen with the highest priority score.

2. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 1 is characterized in that: Said S1 comprises: S11: Real-time environmental image data is collected through a wide-angle camera deployed on the top of the store; S12: Input the environmental image data into the YOLOv7 model to detect the human body bounding boxes of all customers in the image; S13: Based on the human body bounding box, extract the coordinates of the skeleton key points of each customer through the HRNet model; S14: Calculate the angle between the hip joint key point connection vector and the normal direction of the target LED screen according to the hip joint key point connection vector in the skeleton key point coordinates, and output the body orientation angle.

3. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 2 is characterized in that: Said S1 further comprises: S15: Detecting the coordinates of the customer's facial feature points within the human body bounding box area using the RetinaFace model; S16: Calculate the angle between the vector connecting the pupil center and the nose tip in the facial feature point coordinates and the normal direction of the target LED screen, and output the facial orientation angle; S17: Aggregate the body orientation angles and facial orientation angles of all customers to generate customer status data including each customer ID and its corresponding angle value.

4. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 3 is characterized in that: The S2 includes: S21: extracting the facial orientation angles and body orientation angles of all current customers from the customer status data; S22: The linear distance between customers and each LED screen is obtained in real time through a group of UWB positioning base stations pre-installed on the store floor; S23: Calculate the cosine value of the angle between the customer's facial orientation vector and the normal direction of the target LED screen according to the facial orientation angle to generate a facial orientation factor; S24: Calculate the cosine value of the angle between the customer's body orientation vector and the normal direction of the target LED screen according to the body orientation angle to generate a body orientation factor; S25: Calculating a distance attenuation factor based on the spatial straight-line distance by applying an exponential function; S26: Based on the facial orientation factor, the body orientation factor and the distance attenuation factor, the customer's real-time attention weight to the target LED screen is calculated.

5. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 4 is characterized in that: The attention weight is calculated as: Attention weight = (face orientation factor + 0.5 × body orientation factor) × distance decay factor.

6. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 5 is characterized in that: The S3 includes: S31: For the same LED screen, the output attention weights of all customers are accumulated and summed to generate the aggregated attention of the screen; S32: Retrieve pre-stored product heat map data from the store backend system in real time. The product heat map data includes the weight score of each product corresponding to the screen; S33: Multiplying the aggregated attention by the weight score of the corresponding product in the product heat map data to obtain the basic content push priority score of each LED screen; S34: When the difference between the basic content push priority scores of two or more LED screens is less than a preset threshold, the score weight of the screen associated with the high inventory warning level product is increased to generate a final content push priority score.

7. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 6 is characterized in that: The weight score of the screen corresponding to the product is generated by weighted calculation of historical customer stay time, touch interaction frequency and inventory warning level.

8. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 7 is characterized in that: The S4 includes: S41: Filtering the LED screen corresponding to the highest score from the content push priority scores to generate a target LED screen identifier; S42: According to the product ID associated with the target LED screen identifier, searching for multimedia content bound to the product ID in a preset content library, and outputting matching multimedia content; S43: dynamically adapting and rendering the matching multimedia content to the physical size parameters of the target LED screen to generate a content update instruction data packet that can be parsed by the screen; S44: Sending a content update instruction data packet to the hardware device corresponding to the target LED screen identifier via the store local area network to trigger a refresh of the screen display content.

9. The method for intelligently pushing screen content of store LED display cabinets based on image recognition technology according to claim 8 is characterized in that: In S42, when there is no matching content in the preset content library, the default promotional video of the same category of products is automatically called, the backup interactive game program library is activated, and a content missing alarm is sent to the background system.

Images