User target tracking analysis method and system based on AI visual identification

Abstract

The invention relates to the field of visual identification, in particular to a user target tracking analysis method and system based on AI visual identification. The method comprises the following steps: acquiring a real-time user purchase image based on a vending machine camera; performing multi-face positioning detection on a real-time user purchase image, and performing face detail optimization so as to construct a detail-sharpened face image; acquiring a historical monitoring log of the vending machine; performing user identity intelligent perception and user purchase behavior analysis on the detail sharpened face image according to a historical monitoring log of the vending machine to obtain optimized user purchase behavior data; multi-time-point purchase habit mining and user purchase demand analysis are performed on the optimized user purchase behavior data, so that a personalized purchase portrait of the user is constructed; and eyeball position positioning is carried out on the detail-sharpened face image, and continuous inter-frame sight change tracking is carried out to extract a user sight focus moving track. According to the invention, efficient and accurate user target tracking is realized, and the purchase experience of the user is improved.

Description

Technical Field

[0001] The present invention relates to the field of visual recognition, and in particular to a user target tracking and analysis method and system based on AI visual recognition. Background Art

[0002] With the continuous development of science and technology, vending machines, as an efficient and convenient retail method, have gradually become an important part of the modern business environment. Traditional vending machines usually rely on simple product display and payment functions. However, with the increasing personalization and intelligence of consumer needs, modern vending machines have begun to develop in a more intelligent and personalized direction, aiming to provide consumers with a more interactive and precise shopping experience.

[0003] In recent years, research based on face recognition and eye tracking technology has gradually become a key means to enhance the smart retail experience. By capturing the user's facial features in real time and analyzing their eye movement trajectories, it is possible to accurately obtain the user's interests and attention behaviors, and then track and analyze their shopping behaviors in real time. However, traditional vending machine selection item analysis methods usually rely on static data records and user manual input, which not only leads to a lack of accuracy in the analysis results, but also often fails to reflect the user's actual interests and needs, resulting in unsatisfactory shopping recommendation results.

[0004] In this context, a method for tracking and analyzing user vending machine item selection based on face and eye tracking recognition has emerged. By combining face recognition technology and eye tracking technology, this method can track the user's facial features and line of sight changes in real time when the user uses the vending machine, accurately capture the user's shopping intention, and then optimize product recommendations and sales strategies. Compared with traditional methods, the analysis method based on face and eye tracking can better realize personalized recommendations, reduce confusion and time waste in the user selection process, improve the shopping experience, and significantly increase the purchase conversion rate.

[0005] With the continuous improvement of industrial automation and intelligence, vending machines, as an important innovation in the retail industry, have gradually been integrated into modern business and daily life. Therefore, how to use advanced technical means, such as face recognition and eye tracking, to improve the interactivity and accuracy of vending machines has become an important direction of research and application in the field of smart retail. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a user target tracking and analysis method and system based on AI visual recognition to solve at least one of the above technical problems.

[0007] To achieve the above object, the present invention provides a user target tracking and analysis method based on AI visual recognition, comprising the following steps: Step S1: obtaining a real-time user purchase image based on a vending machine camera; performing multi-face location detection on the real-time user purchase image, and optimizing face details, thereby constructing a detail-sharpened face image; Step S2: Obtain the vending machine history monitoring log; perform intelligent user identity perception and user purchase behavior analysis on the detail-sharpened face image according to the vending machine history monitoring log to obtain optimized user purchase behavior data; Step S3: Optimize the user's purchase behavior data to mine the purchase habits at multiple time points and analyze the user's purchase needs, so as to build a personalized purchase profile of the user; Step S4: locating the eyeball position of the detail-sharpened face image and tracking the sight line changes between consecutive frames to extract the movement trajectory of the user's sight line focus; Step S5: performing a sight cross-matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculating the user's sight stay time, thereby obtaining the user's sight stay time; Step S6: predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

[0008] The present invention ensures that the user's facial features can be accurately captured even in the case of unsatisfactory ambient light or complex background through multi-face positioning detection and detail sharpening processing. By image sharpening and noise filtering, the accuracy of facial recognition and eye tracking in the subsequent analysis process is improved, and the influence of false recognition and error is reduced. Through the docking of historical monitoring logs, the system can identify and associate the historical behavior data of the current user, including purchase preferences, frequency, categories, etc. Combined with real-time image analysis, the system intelligently perceives the identity of the current user, and optimizes the data on this basis, thereby providing users with more personalized services. This makes the vending machine not only a simple transaction tool, but also an intelligent shopping companion for users. Combined with historical monitoring logs, user behavior analysis can more accurately understand the shopping habits of each user. By mining the purchase data of each user through intelligent algorithms, the changes in user demand can be more accurately predicted, thereby optimizing the recommendation strategy of goods. By deeply mining the purchase data of users at multiple time points, the system can identify the user's purchase pattern and cycle law. This makes the personalized purchase portrait of each user not only limited to the past purchase history, but also based on the changes in demand in different time periods for real-time optimization, ensuring that the recommended content is closer to the user's immediate needs. The generation of personalized portraits allows vending machines to intelligently adjust product recommendations based on different user characteristics. This personalized recommendation method based on behavioral data, demand analysis and user portraits makes the products recommended by vending machines more targeted, thereby improving product selection efficiency and purchase conversion rate. By processing the detailed sharpened face image, the system can track the changes in the position of the user's eyeballs in real time between consecutive frames to ensure that the user's eye focus is accurately captured. This technology breaks through the technical bottleneck of traditional eye tracking, allowing the system to effectively identify the user's eye movement pattern and accurately track their focus. By tracking eye movements between consecutive frames, the system can identify changes in the user's line of sight and ensure that the eye focus is updated in real time. This technology can provide real-time feedback to the system, thereby improving the user experience and ensuring that the vending machine can dynamically adjust according to the user's current visual focus. By cross-matching the eye focus with the product area, the system can determine in real time which products the user is more interested in. This process can accurately reflect the user's focus and help the vending machine identify the user's interest in the item. As an important indicator, the length of time the eye stays can help the system evaluate the user's purchasing tendency and provide personalized recommendations. By calculating the time the user's gaze stays on a product in real time, the system can display detailed information about the product the user is currently interested in on the touch screen, creating an interactive shopping experience. This not only makes the user feel more convenient and caring, but also improves the user experience and efficiency of the vending machine. Combining the user's gaze stay time and personalized purchase portrait, the system can accurately predict the user's purchase intention.The duration of gaze retention is a key indicator for judging user interest and purchase possibility. Through this data, the vending machine can identify user needs in advance and further improve the purchase conversion rate. After the user's purchase intention is confirmed, the vending machine can accurately locate the products that the user is interested in and perform target correction. Through this intelligent item positioning, the system can dynamically adjust the position and order of product display according to the actual needs of the user, making it easier for users to find and select the products they are interested in. This intelligent reasoning method based on user behavior and intention can significantly increase the frequency of interaction between users and vending machines and increase purchase conversion rates. Through accurate intention prediction and product positioning, the vending machine can push products of interest to users, thereby improving the user's shopping experience.

[0009] In this specification, a user target tracking and analysis system based on AI visual recognition is provided, which is used to execute the user target tracking and analysis method based on AI visual recognition as described above, including: The image processing module is used to obtain real-time user purchase images based on the vending machine camera; perform multi-face positioning detection on the real-time user purchase images, and optimize the face details, thereby constructing a detail-sharpened face image; The identity intelligent perception module is used to obtain the historical monitoring logs of the vending machine; based on the historical monitoring logs of the vending machine, the user identity intelligent perception and user purchase behavior analysis are performed on the detail-sharpened face image to obtain optimized user purchase behavior data; Personalized portrait module, used to optimize user purchase behavior data, multi-time point purchase habit mining and user purchase demand analysis, so as to build a personalized purchase portrait of the user; The gaze tracking module is used to locate the eyeball position of the detail-sharpened face image and track the gaze changes between consecutive frames to extract the movement trajectory of the user's gaze focus; The sight cross matching module is used to perform sight cross matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculate the user's sight stay time, so as to obtain the user's sight stay time; The user intention prediction module is used to predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

[0010] The present invention helps to improve the clarity of the image through detail optimization technology, especially in the case of insufficient light or complex background. The optimized image ensures higher facial feature recognition accuracy and provides a clear basis for subsequent eye tracking and identity perception. In multi-person scenarios, the system can effectively identify and extract the facial features of each user, improve the accuracy and fault tolerance of the system, and work effectively even in dynamic environments. By combining historical behavior with real-time images, the system can intelligently identify users and track their historical behavior to achieve personalized recommendations. For example, the system can provide customized services for users based on their historical purchasing behavior, preferences, purchasing frequency and other information. Through the analysis of user purchasing behavior by deep learning and machine learning algorithms, it is possible to more accurately capture changes in user needs, thereby improving the relevance and accuracy of recommendations and optimizing the shopping experience. By analyzing purchasing behavior at different time points, it is possible to identify and predict the potential needs of users. For example, users may be more inclined to buy specific types of goods (such as holiday gifts, seasonal goods, etc.) in certain specific time periods, and the generation of personalized portraits helps to prepare for product recommendations in advance. By combining the user's purchasing habits, the vending machine can recommend products that are more in line with the user's interests and improve the conversion rate of purchase decisions. By accurately locating the eyeball position, the system can accurately track the user's sight focus and determine which products or areas the user is interested in. This provides basic data for subsequent product recommendations and user intention prediction. By continuously tracking the user's sight changes, the vending machine can adjust the product display in real time according to the user's attention changes, forming a more interactive and personalized shopping experience. By calculating the user's sight dwell time in the product area, the system can dynamically understand the user's purchasing interest and identify high-interest products. This helps the vending machine adjust the display content in real time, push products that the user is interested in, and improve the display effect of the product. Through the cross-matching analysis of sight, the system can not only identify the user's interest points, but also push information about related products or provide promotional reminders based on the dwell time to optimize purchase decisions. Combined with the sight dwell time and user portrait, the system can accurately predict the user's purchase intention and identify the user's purchase needs in advance. The timely prediction of user intention provides real-time data support for subsequent product recommendations and displays. By accurately predicting user intentions and accurately locating items, the vending machine can quickly respond to user needs and improve the purchase conversion rate. At the same time, the system can also optimize the product display order and the recommended information pushed in real time, thereby improving the user experience and the overall efficiency of the vending machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A schematic diagram of the steps of a user target tracking and analysis method based on AI visual recognition according to the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Detailed implementation flow chart of step S3. DETAILED DESCRIPTION

[0012] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0013] The example of this application provides a user target tracking analysis method and system based on AI visual recognition. The execution subject of the user target tracking analysis method and system based on AI visual recognition includes but is not limited to: mechanical equipment, data processing platform, cloud server node, network upload device, etc. equipped with the system can be regarded as the general computing node of this application, and the data processing platform includes but is not limited to: at least one of an audio image management system, an information management system, and a cloud data management system.

[0014] See also Figures 1 to 4 The present invention provides a user target tracking and analysis method based on AI visual recognition, and the user target tracking and analysis method based on AI visual recognition comprises the following steps: Step S1: obtaining a real-time user purchase image based on a vending machine camera; performing multi-face location detection on the real-time user purchase image, and optimizing face details, thereby constructing a detail-sharpened face image; Step S2: Obtain the vending machine history monitoring log; perform intelligent user identity perception and user purchase behavior analysis on the detail-sharpened face image according to the vending machine history monitoring log to obtain optimized user purchase behavior data; Step S3: Optimize the user's purchase behavior data to mine the purchase habits at multiple time points and analyze the user's purchase needs, so as to build a personalized purchase profile of the user; Step S4: locating the eyeball position of the detail-sharpened face image and tracking the sight line changes between consecutive frames to extract the movement trajectory of the user's sight line focus; Step S5: performing a sight cross-matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculating the user's sight stay time, thereby obtaining the user's sight stay time; Step S6: predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

[0015] The present invention ensures that the user's facial features can be accurately captured even in the case of unsatisfactory ambient light or complex background through multi-face positioning detection and detail sharpening processing. By image sharpening and noise filtering, the accuracy of facial recognition and eye tracking in the subsequent analysis process is improved, and the influence of false recognition and error is reduced. Through the docking of historical monitoring logs, the system can identify and associate the historical behavior data of the current user, including purchase preferences, frequency, categories, etc. Combined with real-time image analysis, the system intelligently perceives the identity of the current user, and optimizes the data on this basis, thereby providing users with more personalized services. This makes the vending machine not only a simple transaction tool, but also an intelligent shopping companion for users. Combined with historical monitoring logs, user behavior analysis can more accurately understand the shopping habits of each user. By mining the purchase data of each user through intelligent algorithms, the changes in user demand can be more accurately predicted, thereby optimizing the recommendation strategy of goods. By deeply mining the purchase data of users at multiple time points, the system can identify the user's purchase pattern and cycle law. This makes the personalized purchase portrait of each user not only limited to the past purchase history, but also based on the changes in demand in different time periods for real-time optimization, ensuring that the recommended content is closer to the user's immediate needs. The generation of personalized portraits allows vending machines to intelligently adjust product recommendations based on different user characteristics. This personalized recommendation method based on behavioral data, demand analysis and user portraits makes the products recommended by vending machines more targeted, thereby improving product selection efficiency and purchase conversion rate. By processing the detailed sharpened face image, the system can track the changes in the position of the user's eyeballs in real time between consecutive frames to ensure that the user's eye focus is accurately captured. This technology breaks through the technical bottleneck of traditional eye tracking, allowing the system to effectively identify the user's eye movement pattern and accurately track their focus. By tracking eye movements between consecutive frames, the system can identify changes in the user's line of sight and ensure that the eye focus is updated in real time. This technology can provide real-time feedback to the system, thereby improving the user experience and ensuring that the vending machine can dynamically adjust according to the user's current visual focus. By cross-matching the eye focus with the product area, the system can determine in real time which products the user is more interested in. This process can accurately reflect the user's focus and help the vending machine identify the user's interest in the item. As an important indicator, the length of time the eye stays can help the system evaluate the user's purchasing tendency and provide personalized recommendations. By calculating the time the user's gaze stays on a product in real time, the system can display detailed information about the product the user is currently interested in on the touch screen, creating an interactive shopping experience. This not only makes the user feel more convenient and caring, but also improves the user experience and efficiency of the vending machine. Combining the user's gaze stay time and personalized purchase portrait, the system can accurately predict the user's purchase intention.The duration of gaze retention is a key indicator for judging user interest and purchase possibility. Through this data, the vending machine can identify user needs in advance and further improve the purchase conversion rate. After the user's purchase intention is confirmed, the vending machine can accurately locate the products that the user is interested in and perform target correction. Through this intelligent item positioning, the system can dynamically adjust the position and order of product display according to the actual needs of the user, making it easier for users to find and select the products they are interested in. This intelligent reasoning method based on user behavior and intention can significantly increase the frequency of interaction between users and vending machines and increase purchase conversion rates. Through accurate intention prediction and product positioning, the vending machine can push products of interest to users, thereby improving the user's shopping experience.

[0016] In the embodiment of the present invention, refer to Figure 1 , is a schematic flow chart of the steps of a user target tracking and analysis method based on AI visual recognition of the present invention. In this example, the steps of the user target tracking and analysis method based on AI visual recognition include: Step S1: obtaining a real-time user purchase image based on a vending machine camera; performing multi-face location detection on the real-time user purchase image, and optimizing face details, thereby constructing a detail-sharpened face image; In this embodiment, ensure that the camera of the vending machine is installed in a suitable position to capture the user's facial image. The camera should have a high resolution (such as 1080p) and an appropriate viewing angle to cover the entire area in front of the vending machine, ensuring that the user can be clearly captured when purchasing. Debug the camera to ensure that it can work effectively under different lighting conditions (such as natural light, artificial light). Preliminary tests can be performed to check the clarity and stability of the image to ensure that there is no blur or jitter during the real-time capture process. Use the camera software or hardware interface to set the real-time image capture function. The capture frame rate (such as 30 frames / second) can be set to ensure smooth images and avoid missing user facial information due to frame loss. During the capture process, record the timestamp of each image to facilitate subsequent data processing and analysis. Save the captured image in a temporary storage area for subsequent processing. Implement image quality monitoring to ensure that the captured image meets the standard. Quality detection parameters such as brightness, contrast and clarity can be set to automatically filter out low-quality images. For images that do not meet the standards, they can be recaptured or discarded to ensure the quality of data for subsequent processing. Image processing software can be used to automate monitoring and reduce manual intervention. Use modern face detection algorithms (such as Haar Cascades, Dlib, or MTCNN) to detect multiple faces in real-time user purchase images. These algorithms can identify multiple facial features in the image and return the location coordinates of each face (such as the upper left and lower right corners of a rectangular box). To improve the accuracy of detection, the image can be preprocessed, such as converting to grayscale images, adjusting brightness and contrast, etc. This helps reduce background noise and improve the performance of the face detection algorithm. After successfully locating the face, each face is optimized for detail. Image sharpening techniques (such as Unsharp Masking or Laplacian sharpening) can be used to enhance the details of the face so that the facial features are more clearly distinguishable. During the sharpening process, set appropriate parameters, such as sharpening strength and radius, to ensure that the facial details are enhanced without introducing too much noise or artifacts. These parameters can be adjusted through experiments to find the best sharpening effect. Save the face image with detail optimization as a new image file to ensure that each user's face image is recorded and associated with the user's purchase behavior. Metadata can be attached to each detailed face image, such as user ID, timestamp, and detection results. Ensure the image storage format (such as JPEG or PNG) and compression quality for subsequent data processing and analysis. The image save path should be easily accessible for subsequent user identification and behavior analysis.

[0017] Step S2: Obtain the vending machine history monitoring log; perform intelligent user identity perception and user purchase behavior analysis on the detail-sharpened face image according to the vending machine history monitoring log to obtain optimized user purchase behavior data; In this embodiment, confirming that the monitoring log system of the vending machine is operating normally includes ensuring the accessibility of the database and the integrity of the data storage. The monitoring log usually contains user interaction records, such as purchase time, purchased items, user ID, payment method and other information. Check the format and content of the log through a database management tool (such as SQL management tool) to ensure that each record contains key fields, including timestamp, user ID, product ID and status information. As needed, query conditions can be set to extract log data within a specific time range. Use appropriate query statements to extract the required monitoring log data from the database. Filter conditions can be set, such as filtering by date range, user ID or specific product to limit the amount of data extracted and increase processing efficiency. Export the extracted data into a structured file format (such as CSV or Excel) for subsequent analysis. Ensure that all relevant information is retained during the export process, and record the time and source of data extraction. Perform a preliminary data quality check on the extracted historical monitoring logs to identify and process missing values, duplicate records and abnormal data. Statistical methods, such as descriptive statistical analysis, can be used to analyze the basic trends of user behavior. Record the basic statistical information of the log, including the total number of users, total number of purchases, sales of different products, etc., to provide a basis for subsequent data analysis. Combine the extracted detail-sharpened face image with the historical monitoring log, and use machine learning algorithms (such as support vector machine, random forest, etc.) to identify the user. The key to intelligent perception of user identity is to associate the features of the face image with the user ID in the monitoring log. Extract the feature points in the detail-sharpened face image, construct the feature vector of each user, and compare it with the user behavior in the historical monitoring log to identify the user's identity. This process may require the use of a pre-trained model to ensure the accuracy of recognition. Analyze the purchase behavior of the identified user identity and extract the user's purchasing habits and preferences. Cluster analysis methods (such as K-means or hierarchical clustering) can be used to divide users into different groups to identify the purchasing behavior characteristics of different user groups. Analyze important indicators such as user purchase frequency, average consumption amount, and purchase time interval to deeply understand user behavior. Visualization tools (such as Tableau or Matplotlib) can be combined to intuitively display the user's purchasing behavior pattern. Integrate the identified user identity and its purchasing behavior data into an optimized user purchase behavior dataset. Ensure that each user's behavior data is associated with identity information, purchasing habits, preferences, etc. for subsequent analysis. Save the optimized user purchase behavior data to the database to ensure data integrity and traceability. Record key parameters and results during the analysis process to provide data support for subsequent market strategies and personalized recommendations.

[0018] Step S3: Optimize the user's purchase behavior data to mine the purchase habits at multiple time points and analyze the user's purchase needs, so as to build a personalized purchase profile of the user; In this embodiment, the optimized user purchase behavior data is cleaned and sorted for subsequent analysis. Ensure that the data includes purchase records at multiple time points, such as purchase time, item ID, quantity, and user ID. Format the timestamp to ensure that all time information in the data is consistent (such as unified as UTC time), and fill or remove missing values ​​to ensure data integrity. Convert the user's purchase behavior data into a time series to analyze their purchasing habits and trends. The time window method (such as day, week, month) can be used to count the number of purchases of each user at different time points. Use sliding window technology to analyze the changes in user purchase frequency in different time periods, and identify peak periods (such as a significant increase in the number of purchases in a specific time period) and trough periods (a significant decrease in the number of purchases). This helps to understand the user's purchase cycle and preferences. Use a clustering algorithm (such as K-means or DBSCAN) to group user purchase behaviors to identify user groups with similar purchasing habits. Set appropriate clustering parameters (such as the number of clusters k) and determine the optimal clustering results through experiments. Analyze the user characteristics of different clusters to identify the main purchasing habits and preferences of each group. This information will provide basic data for the construction of personalized user portraits. Based on the results of multi-time point purchase habit analysis, extract user demand features, such as purchase frequency, purchase amount, purchase time preference, etc. Quantify these features to form the user's demand feature vector. Combine the user's purchase history to identify the user's potential needs. For example, analyze the user's purchase tendency for a certain type of item in a specific time period in order to predict their future purchase needs. Apply association rule mining technology (such as Apriori algorithm) to analyze the user's purchase behavior to identify potential associations between items. For example, analyze which products are often purchased together to form the rule of "if you buy A, you may buy B". Record the support and confidence of each association rule to evaluate its strength and reliability. Through these rules, you can recommend products related to your purchase habits to users and improve the effect of personalized services. Integrate the data obtained through multi-time point purchase habit mining and demand analysis to build a personalized user purchase portrait. The portrait should include multiple dimensions such as the user's basic information, purchase preferences, demand characteristics, and product categories of interest. When building a user portrait, ensure that the data is dynamically updated and regularly analyze the user's new purchase behavior to adapt to their changing needs and preferences. You can set a period (such as monthly) to update and optimize user portraits.

[0019] Step S4: locating the eyeball position of the detail-sharpened face image and tracking the sight line changes between consecutive frames to extract the movement trajectory of the user's sight line focus; In this embodiment, before starting to locate the eyeball position, first ensure that the quality of the face image with sharpened details is good. The preprocessing step includes converting the image to a grayscale image to reduce the computational complexity, while enhancing the contrast of the eye features to make the eyeball more obvious in the image. Gaussian blur technology is applied to remove noise in the image to reduce interference with the eyeball positioning algorithm. The blur radius (such as 3x3 or 5x5) can be set to achieve the best effect. Select a suitable eyeball detection algorithm, such as Haar cascade classifier or Dlib's facial feature detection method. Haar cascade classifiers have been trained extensively to quickly and effectively locate the eye area, while Dlib provides detection of 68 facial feature points, which can more finely identify the eyeball position. When performing eyeball detection, ensure that the parameters of the algorithm have been optimized, such as setting an appropriate detection ratio and neighborhood value to improve the accuracy and robustness of the detection. Apply the selected eyeball detection algorithm to the face image with sharpened details to identify the position of the user's eyeballs. The algorithm will return a coordinate box for each eye, usually including the coordinates of the upper left corner and the lower right corner. Record the position information of each eyeball, including coordinates and detection confidence, for subsequent analysis. A confidence threshold (such as 0.7) can be set to ensure that only high-confidence detection results are recorded. Obtain real-time video streams from the camera of the vending machine to ensure that the user's facial image can be captured. Set an appropriate frame rate (such as 30 frames per second) to ensure continuity and smoothness, and ensure the clarity and stability of each frame. Each frame of the image undergoes the same preprocessing steps to ensure the accuracy of eyeball position detection and maintain consistency in image quality. Use algorithms such as optical flow or Kalman filter to track the eyeball position between consecutive frames. The optical flow method can infer the movement trajectory of the line of sight based on the changes in the eyeball position in consecutive frames. When implementing line of sight tracking, it is necessary to set an appropriate window size (such as 5x5) to smooth the movement of the eyeball and reduce the error caused by slight jitter. Record the position information of the eyeball in each frame and convert it into a trajectory of the movement of the line of sight focus. These trajectory data contain the eyeball coordinates and timestamps at each time point for subsequent analysis. Analyze the gaze movement trajectory, calculate the gaze movement speed and direction, and identify the user's gaze point. Statistical methods can be used to analyze the user's gaze duration on different objects to understand the user's focus.

[0020] Step S5: performing a sight cross-matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculating the user's sight stay time, thereby obtaining the user's sight stay time; In this embodiment, the preset area information of each item on the vending machine is prepared. This includes the coordinates of each item on the screen (such as the upper left corner and lower right corner of the rectangular box), size and item name, etc. Ensure that this information is accurate for subsequent matching analysis. Store the item area information in a structured database for real-time query and call. In the data table, the information of each item should include item ID, item name, area coordinates and other related attributes. From the user's sight focus movement trajectory extracted in the previous step, obtain the eyeball coordinates and timestamp at each time point. Arrange these data to ensure that the data format is consistent for subsequent analysis. Assign a timestamp to each sight point to form an ordered time series data set. Each data point should contain time, X coordinate and Y coordinate for accurate subsequent analysis. Use a spatial matching algorithm (such as rectangular intersection detection) to cross-match the user's sight trajectory with the item area. Set a suitable matching threshold (such as 5 pixels) to ensure that the overlap between the sight point and the item area can effectively identify the user's focus point. For each sight point, determine whether it falls within any item area. If it falls within the area, record the ID and corresponding timestamp of the item. In this way, a list of items that the user's gaze focuses on can be generated for subsequent analysis. Set a method for calculating the gaze dwell time. The timestamp difference method can be used to calculate the dwell time by recording the time points when the user's gaze enters and leaves a certain object area. Set a time threshold (such as 100 milliseconds) to ensure that only the time the user actually stays is recorded to avoid short-term gaze deviations. Design a state machine to track the state changes of the user's gaze (such as "entering an object" and "leaving an object"). This can effectively manage the entry and exit events of the gaze and ensure accurate calculation of the dwell time. Traverse the user's gaze movement trajectory, identify the time point when the gaze enters a certain object area, and record the timestamp. When the gaze leaves the area, record the timestamp again. By calculating the time difference between the two, the user's dwell time on the object is obtained. Summarize the dwell time of all objects and calculate the average dwell time and total dwell time of each object. A data structure can be set to store the dwell time information of each object, including the object ID, total dwell time, and average dwell time.

[0021] Step S6: predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

[0022] In this embodiment, before predicting the user's purchase intention, the user's sight dwell time is first integrated with the personalized purchase portrait. The sight dwell time provides the user's attention to each item, while the personalized purchase portrait contains the user's historical purchase behavior, preferences and demand characteristics. Extract key features, such as dwell time, user's purchase frequency, purchase amount, and concerned item category, to form a feature vector. Set feature weights to ensure the importance of dwell time in prediction. The correlation between dwell time and purchase intention can be determined through previous experimental analysis. Select a suitable machine learning model for purchase intention prediction, such as logistic regression, random forest, or gradient boosting tree (GBT). These models can handle nonlinear relationships between different features and provide probabilistic predictions for users' purchase intentions. Use historical user data for model training to ensure the correlation between input features and output labels (purchase or not). The cross-validation method can be used to evaluate model performance to ensure the generalization ability and prediction accuracy of the model. Input the integrated feature vector into the trained model to predict the user's purchase intention. The model will output a probability value indicating the strength of the user's purchase intention for a specific item. Based on the set threshold (e.g., 0.5), determine whether the user has a strong purchase intention, and then determine whether the item should be marked as a target of interest. Update the regional information of the items in the vending machine based on the user's purchase intention prediction results. This includes the coordinates, size, and status of the items, etc., to ensure that the items of interest to the user can be clearly displayed on the vending machine interface. Store the location information of the relevant items in the database for real-time call and update. During the update process, ensure that the item status (such as availability, inventory, etc.) matches the user's prediction results. For items predicted as targets of interest, use a positioning algorithm (such as coordinate-based matching) to highlight them on the vending machine interface. Graphics rendering technology can be used to dynamically update the display status of items on the user interface. Set the parameters of the highlight display, such as border color, transparency, and flashing frequency, to attract user attention and improve user experience. These parameters can be adjusted through user feedback to achieve the best effect. Confirm the final user tracking target and record relevant information, including user ID, predicted purchase intention items, item location, and timestamp. Store this information in the database for subsequent behavior analysis and personalized services. The database structure can be set to ensure data traceability and integrity.

[0023] In this embodiment, refer to Figure 2 , is a flowchart of detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Step S11: Acquire real-time user purchase images based on the vending machine camera; Step S12: performing multi-face location detection on the real-time user purchase image and marking multiple user face image frames; Step S13: calculating the face positions of multiple user face image frames one by one to obtain the distance of each face; Step S14: performing the closest distance recognition relative to the vending machine according to the distance of each face, and extracting the closest face image frame; Step S15: Optimizing the facial details of the nearest facial image frame, thereby constructing a detail-sharpened facial image.

[0024] In this embodiment, ensure that the camera on the vending machine can work properly and has sufficient resolution and frame rate. It is recommended to use a 1080p high-definition camera that can capture real-time video at a rate of 30 frames per second. This setting can ensure the clarity and smoothness of the image and facilitate subsequent processing. Configure the camera's viewing angle so that it can cover the interaction area between the user and the vending machine. The angle and height of the camera can be adjusted to ensure that the user's facial image can be captured in any situation. Start the camera to capture the real-time video stream. Ensure the stability of the video stream to avoid image quality degradation due to light changes or motion blur. Use image processing software (such as OpenCV) to perform real-time analysis on the captured video stream and extract each frame of the image regularly. It can be set to extract one frame every 100 milliseconds to obtain sufficient user purchase behavior data. Save the extracted image data to a database or local storage to ensure data integrity and security. It can be stored in JPEG or PNG format to ensure image clarity. Record the timestamp and related metadata (such as the number of users, purchased items, etc.) of each frame of the image to provide necessary context information for subsequent analysis. Select a suitable multi-face detection algorithm, such as Haar cascade classifier, Dlib, or YOLO (You Only Look Once). YOLOv3 or YOLOv4 models are recommended, which can achieve efficient and accurate face detection and are suitable for real-time processing. Before performing face detection, ensure that the algorithm model is properly trained and optimized to improve detection accuracy and speed. Before performing face detection, preprocess the acquired images, including resizing, grayscale, and denoising. Reduce the amount of calculation and speed up processing by scaling the images to a uniform size (such as 640x480 pixels). Use Gaussian blur or median filtering to remove noise in the image to ensure more accurate face detection. Process each frame of the image using the selected face detection algorithm to detect the locations of all faces in the image. The algorithm returns the coordinate box (bounding box) of each detected face. Draw a marker box on the image to identify the face area of ​​each user, and save the marked image for subsequent use. You can use different colored boxes to distinguish different faces. Before calculating the distance, you need to know the parameters of the camera, such as focal length, sensor size, etc. These parameters are crucial for the subsequent distance estimation. This information can be obtained through camera calibration. Set a known reference object (such as a marker on a vending machine) for relative distance calculation. Perform center point calculation for each marked face frame to determine the position of each face in the image. The center point coordinates can be obtained by calculating the average of the coordinates of the upper left corner and the lower right corner of the frame. Use the principle of similar triangles to estimate the distance based on the actual size of the face frame and the size relationship in the image.Assuming that the average width of a face is 15 cm, the ratio can be calculated based on the pixel width of the face frame in the image and the camera's viewing angle. The distance of each face is recorded in the database, including the coordinates of the face frame and the calculated distance value. Ensure the accuracy of the data for subsequent analysis and processing. By analyzing the distance data, the relative position between the user and the vending machine is identified, and the nearest face is prepared for extraction. From the recorded face distance data, the face closest to the vending machine is filtered out. The data corresponding to the minimum value can be found by traversing all recorded distance values. To ensure accuracy, a threshold is set to ensure that the selected nearest face is within a reasonable distance range (such as within 1 meter) to exclude unreasonable selection due to errors. According to the coordinate information of the nearest face identified, the nearest face image frame is extracted from the marked image. The slicing operation can be used to obtain the image portion of the face frame. Save the extracted nearest face image for subsequent detail optimization processing. The image frame coordinates and distance information of the nearest face are recorded in the database for subsequent analysis. At the same time, the user's relevant behavioral data (such as purchase intention) is marked to provide support for subsequent system decisions. Select an appropriate image processing algorithm for detail optimization. Common algorithms include histogram equalization, CLAHE (Contrast Limited Adaptive Histogram Equalization), and sharpening filters. It is recommended to use sharpening filter algorithms, such as Laplacian operator or Unsharp Masking, to enhance the details and clarity of the face image. Before performing detail optimization, preprocess the extracted nearest face image, including adjusting the image size and color space conversion (such as from RGB to YUV). Adjust the image contrast and brightness to ensure that it will not be overexposed or dim during detail optimization. Apply the selected sharpening algorithm to process the nearest face image to enhance facial details. For the Unsharp Masking method, the best effect can be achieved by setting an appropriate blur radius and enhancement factor. After processing, observe the clarity and detail changes of the image to ensure that the optimized image can clearly show the facial features.

[0025] In this embodiment, the specific steps of step S15 are: Calculate the grayscale histogram of the nearest face image frame and extract the image grayscale histogram; Dynamically stretch the image grayscale histogram to obtain an adaptively stretched grayscale image; Perform image fusion reconstruction on the nearest face image frame according to the adaptively stretched grayscale image to construct a brightness optimized face image; Perform high-frequency noise detection on brightness-optimized face images and mark all image noise points; Calculate the pixel values ​​of all the image noise points; Performing adaptive filtering and noise reduction based on the pixel values ​​to obtain a filtered optimized face image; The face contour details of the filtered optimized face image are sharpened to construct a detail-sharpened face image.

[0026] In this embodiment, the extracted recent face image frame is converted from the RGB color space to a grayscale image. This step is achieved by calculating the weighted average of each pixel, and the commonly used weighted formula is: Gray = 0.299*R + 0.587*G+ 0.114*B. This can help highlight the brightness information in the image and facilitate subsequent histogram calculations. Select appropriate image processing tools (such as OpenCV or PIL) for image conversion to ensure that the converted grayscale image retains the details of the original image. A grayscale histogram is generated using a method for calculating the number of pixels at each grayscale level in the image. Each value in the grayscale value range from 0 to 255 is used as the horizontal axis, and the vertical axis represents the number of pixels corresponding to the grayscale value. The frequency of each grayscale level of the histogram is recorded for subsequent dynamic stretching processing. These frequency data can be stored using an array or dictionary structure for quick access and processing. The calculated grayscale histogram data is saved in a database or local file for subsequent analysis and processing. At the same time, a histogram graph is drawn to intuitively display the distribution of different grayscale levels. Analyze the shape and distribution of the histogram to identify the brightness and contrast characteristics of the image, which provides a basis for subsequent dynamic stretching. According to the calculated grayscale histogram, determine the minimum and maximum grayscale values ​​(Min and Max) of the image. These values ​​are used in dynamic stretching to improve the contrast of the image. Calculate the dynamic range of the original image, that is, Range = Max - Min, to ensure that the dynamic range during the processing can cover the entire grayscale level. Use the linear stretching formula to adaptively adjust the grayscale value, the formula is: NewGray = (OriginalGray- Min) * (255 / Range). This will map the grayscale values ​​of the original image to the new range and enhance the contrast of the image. Apply this formula to each pixel to generate an adaptively stretched grayscale image to ensure that the details in the image are preserved during the stretching process. Evaluate the effect of the adaptively stretched grayscale image and check whether the contrast and brightness of the image are significantly improved. Use visualization tools (such as Matplotlib) to show the image comparison before and after stretching to ensure that the stretching effect is as expected. Fuse the adaptively stretched grayscale image with the original nearest face image frame. The purpose of this step is to combine the details of the original image with the contrast of the stretched image to achieve brightness optimization. Choose a suitable fusion method, such as weighted average or multi-resolution fusion. The weighted average method obtains the final image by weighted summing each pixel of the two images. Perform the fusion operation on each pixel using the formula FusedPixel = α * OriginalPixel + (1 - α) * AdjustedPixel, where α is the fusion weight (such as 0.7), and the optimal value can be determined experimentally. By adjusting the α value, experiment with different fusion effects to ensure that the brightness and details of the final image are optimal.Evaluate the brightness-optimized face image after fusion and observe the improvement in details and contrast. The fusion effect can be quantified by image quality evaluation indicators such as PSNR and SSIM. Save the optimized image for subsequent noise detection and processing. Select a suitable noise detection algorithm, such as a high-pass filter or an edge detection algorithm. The high-pass filter can effectively identify high-frequency components in the image, i.e., noise. You can use the Laplacian operator or the Sobel operator for noise detection to determine the obvious noise areas in the image. Apply the selected noise detection algorithm to the brightness-optimized face image to generate a noisy image. Extract the high-frequency part and mark the noise area. You can set a threshold to identify pixels with an intensity exceeding the threshold and mark these noise areas on the original image. Save the marked noise areas to the database and record the coordinates and intensity information of each noise point for subsequent analysis. At the same time, draw a noise distribution map to help visualize the distribution of noise points. Analyze the distribution characteristics of noise points and identify the areas where noise points are concentrated, which provides a basis for subsequent noise reduction processing. Traverse the marked noise areas and extract the pixel value of each noise point. These pixel values ​​can be stored in arrays or lists to facilitate subsequent calculations. Record the specific location (row and column coordinates) of each noise point and its corresponding grayscale value for subsequent analysis and processing. Calculate the statistical characteristics of the extracted noise pixel values, including the mean, standard deviation, and maximum and minimum values. These statistical information can help understand the intensity characteristics of the noise points. By analyzing the statistical data, the distribution and intensity range of the noise points can be identified, providing a basis for adaptive filtering noise reduction. The noise pixel values ​​and their statistical characteristics are recorded in the database for subsequent noise reduction processing and model optimization. Summarize the noise information and form a report to ensure that the data is complete and easy to analyze later. Select an adaptive filtering algorithm, such as adaptive median filtering, adaptive Gaussian filtering, etc. Adaptive median filtering can effectively remove noise points while retaining edge information, and is suitable for processing face images. Determine the window size and adaptive parameters of the filter, and adjust these parameters through experiments to obtain the best noise reduction effect. Apply the adaptive filter to the brightness optimized face image, and dynamically adjust the filter parameters according to the surrounding environment of each pixel. For each pixel, calculate the median value in its neighborhood, and determine whether to replace the original pixel value based on the median value. Continue to adjust the filter parameters to ensure that the facial details are preserved as much as possible while reducing noise. Evaluate the effect of filtering and optimizing the face image, and check the quality of the denoised image through visual and quantitative indicators such as noise ratio and clarity. Save the processed filtered optimized face image to the database to prepare for the subsequent contour detail sharpening. Select a suitable detail sharpening algorithm, such as Unsharp Masking or Laplacian sharpening. The Unsharp Masking method can effectively highlight the details of the image by enhancing the high-frequency components. Determine the sharpening parameters, including blur radius and enhancement factor, and optimize these parameters through experiments to achieve the best effect.Apply a sharpening algorithm to the filtered optimized face image, and calculate the enhancement by comparing the original image with the blurred image and superimpose it on the original image. For UnsharpMasking, use the formula SharpenedImage = OriginalImage + k * (OriginalImage -BlurredImage), where k is the enhancement factor. Experiment with different sharpening effects by adjusting the k value to ensure that the facial details in the final image are significantly improved. Evaluate the image quality after detail sharpening to observe the clarity of facial details and the sharpness of edges. The effect can be judged by visual inspection and image quality assessment indicators such as PSNR and SSIM. Save the face image after detail sharpening to ensure good image quality for subsequent recognition and analysis.

[0027] In this embodiment, refer to Figure 3 , is a flowchart of detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Step S21: Obtaining the vending machine historical monitoring log; Step S22: performing facial geometric feature point recognition on the detail-sharpened facial image, and marking a plurality of facial geometric feature points; Step S23: Based on the historical monitoring log of the vending machine, intelligent user identity perception is performed on multiple facial geometric feature points to extract accurate user identity information; Step S24: Analyze the user's purchase behavior based on the user's accurate identity information and extract all the user's purchase behavior data; Step S25: Eliminate abnormal data from all user purchase behavior data to obtain optimized user purchase behavior data.

[0028] In this embodiment, it is confirmed that the monitoring log storage system of the vending machine is operating normally. The monitoring log is usually stored in a local database or cloud service, and contains information such as user interaction, purchase records and device status. Ensure the access rights and reading speed of the database to extract data efficiently. Check the format and content of the log to confirm that it contains necessary fields, such as timestamp, user ID, purchased products, purchase quantity, payment status, etc. Ensure the integrity of the fields for subsequent analysis. Use a suitable database query tool (such as SQL) to extract data from the monitoring log. You can set query conditions such as time range, user ID, etc. to limit the amount of data extracted and improve processing efficiency. Export the extracted data as a structured file (such as CSV or Excel) to facilitate subsequent analysis and processing. Ensure that all relevant information is retained during the export process. Perform a preliminary analysis of the historical monitoring logs obtained to check the integrity and consistency of the data. Identify missing values, outliers or entries with irregular formats to ensure data quality. Record basic statistical information of the log, including the total number of users, total number of purchases, sales of different products, etc., to provide basic data for subsequent analysis. Select a suitable facial geometric feature point detection algorithm, such as the 68 feature point detection model in the Dlib library or the Haar cascade classifier in OpenCV. These algorithms can efficiently and accurately identify the key feature points of the face, including eyes, nose, mouth, etc. Make sure the algorithm is well trained and can handle facial images under different lighting and angles. Input the face image with sharpened details into the selected feature point detection algorithm for feature point recognition. Depending on the requirements of the algorithm, the image may need to be preprocessed (such as resizing, grayscale conversion, etc.). The recognition algorithm generates the coordinate information of the feature points, records the location of each feature point, and marks these points on the image. You can use different colors of marking points to distinguish different types of feature points. Save the marked image and feature point coordinate information to the database to ensure traceability for subsequent analysis. Record the feature point information of each user image, including the type of feature point, coordinate location, etc. Analyze the distribution and stability of the feature points to ensure the accuracy of the detection results and provide a basis for subsequent user identification. Integrate the user information in the historical monitoring log with the identified facial geometric feature point data. Combine the two data sources through user ID or other unique identifier to form a comprehensive data set containing user characteristics and purchase records. Ensure the accuracy of the integrated data to avoid mismatches caused by missing or inconsistent user information. Apply machine learning algorithms (such as support vector machines or neural networks) to train and identify users' geometric feature points. These algorithms can intelligently determine the identity of users based on the location and distribution pattern of feature points. During the training process, use historical data sets to train the model to ensure that the model can accurately extract user identity information from feature points.Through the trained model, the identity of each identified feature point is judged, and the result is compared with the user information in the historical monitoring log to extract the accurate user identity information. Ensure the accuracy and reliability of the identity extraction process, record the identity information of each user and its corresponding purchase behavior data, and provide support for subsequent analysis. According to the extracted accurate identity information of the user, filter out all purchase records related to the user from the historical monitoring log. Ensure that the extracted data includes purchase time, product type, purchase quantity, payment status, etc. Organize the extracted data into a structured format (such as a data frame or table) for subsequent analysis. Establish a user behavior analysis model to analyze the user's purchase pattern, frequency, and preference. You can use cluster analysis, association rule mining and other methods to identify the user's purchase behavior characteristics. Perform statistical analysis on the user's purchase behavior, calculate important indicators such as the number of purchases, average consumption amount, and purchase interval time, so as to deeply understand the user behavior. Save the results of the user purchase behavior analysis in the database, and record the behavior characteristics and analysis results of each user. At the same time, use visualization tools (such as Matplotlib or Tableau) to draw charts of user purchase behavior to intuitively display user consumption trends and habits. Through the analysis results, identify potential market opportunities and user needs, and provide data support for subsequent market strategies. Select appropriate outlier detection methods, such as the Z-score method, IQR (Interquartile Range) method, or anomaly detection algorithms based on machine learning. These methods can effectively identify and eliminate data that deviates significantly from normal purchase behavior. Determine the detection threshold. For example, for the Z-score method, the threshold is often set to 3, indicating that values ​​exceeding 3 standard deviations are considered abnormal. Apply the selected outlier detection method to all user purchase behavior data to identify abnormal purchase records. For each user's purchase data, calculate its mean and standard deviation to determine whether there are outliers. Mark all detected abnormal data and record its characteristic information (such as purchase amount, time, etc.) for subsequent analysis. Eliminate data based on the identified outliers to ensure the accuracy and reliability of the end-user purchase behavior data. Save the eliminated data as a new data set to avoid affecting subsequent analysis. Record the eliminated abnormal data and its reasons to provide a basis for subsequent data review and optimization. Ensure that the optimized dataset can better reflect users' real purchasing behavior and preferences.

[0029] In this embodiment, refer to Figure 4 , is a flowchart of detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Step S31: performing purchase timestamp identification on the optimized user purchase behavior data to obtain the time point of each purchase behavior; Step S32: Evolving the periodic purchase behavior based on the time point of each purchase behavior to generate the user's periodic purchase pattern; Step S33: mining the multi-time point purchase habits of the user's periodic purchase rules, thereby obtaining the multi-time point purchase habits of the user; Step S34: Analyze the optimized user purchase behavior data to obtain the items purchased each time; Step S35: Calculate the purchase frequency of each item purchased, thereby generating the purchase frequency of each item; Step S36: Analyze the user's purchase demand according to the purchase frequency of each item to obtain the user's personalized purchase demand; Step S37: Perform user profile fitting based on the user's multi-time point purchasing habits and the user's personalized purchasing needs, thereby constructing a personalized purchasing profile of the user.

[0030] In this embodiment, it is confirmed whether the optimized user purchase behavior data contains purchase timestamp information. The purchase timestamp should be a unique identifier for each purchase behavior, usually stored in the format of date and time (such as YYYY-MM-DD HH:MM:SS). Check the integrity of the data to ensure that each record has a corresponding timestamp to avoid affecting subsequent analysis due to missing values. Extract the timestamp of each purchase behavior from the optimized user purchase behavior data. Data processing tools (such as Pandas) can be used to filter and extract the data set to generate a list containing all purchase time points. During the extraction process, ensure that the format of the timestamp is consistent, and perform data format conversion when necessary for subsequent analysis. Record the extracted purchase timestamp in a new data table and associate it with other information of each purchase behavior (such as user ID, purchased items, etc.). Ensure the traceability and retrieval convenience of the data. Perform a preliminary analysis of the timestamp, record the earliest and latest purchase times, understand the time range of the data, and provide a basis for the analysis of the evolution of subsequent periodic purchase behavior. Use the extracted purchase timestamp data to analyze whether the user's purchase behavior has periodic characteristics. Frequency analysis and time series analysis methods can be used to identify the cyclical patterns of user purchase behavior. Set the standard of the cycle (such as day, week, month), and calculate the number of purchases within the set cycle. The sliding window method can be used to analyze the purchase frequency in each time period. Use time series models (such as ARIMA or seasonal decomposition) to model the user's purchase time points and identify potential cyclical changes. These models can help capture trends and seasonal changes over time. Through model fitting, generate the user's cyclical purchase pattern and record the characteristics of purchase behavior in different cycles, such as peak and trough periods. Organize the generated user's cyclical purchase pattern into a report, including the purchase frequency, time period and related statistical data of each cycle. The analysis results can help identify the user's purchase tendency in different time periods. Save the analysis results to the database to provide data support for subsequent multi-time point purchase habit mining. Associate the user's cyclical purchase pattern with the purchase habit and extract the purchase behavior data at multiple time points. The purchase behavior can be classified by day, week, and month by grouping the timestamp. Count the purchase frequency and purchase behavior at each time point to identify the user's purchase pattern in different time periods. Use clustering algorithms (such as K-means or DBSCAN) to analyze multi-time point purchase data and identify the characteristics of users' purchasing habits. Through cluster analysis, users' purchasing behaviors can be divided into different groups to help understand their purchasing habits. Record the purchasing behavior characteristics of each user at a specific time point to form a multi-time point purchasing habit model of the user. Save the identified multi-time point purchasing habits of users in the database to ensure the integrity and retrievability of the data. At the same time, data visualization tools (such as Matplotlib) can be used to show the changes in users' purchasing behaviors at different time points.Analyze and summarize the user's purchasing habits at various time points to provide basic data for subsequent user portrait fitting. Extract the item information corresponding to each purchase behavior from the optimized user purchase behavior data. Ensure that each record contains the name, category, quantity and other information of the purchased items. Use data processing tools (such as Pandas) to filter and organize the data to generate a list of all purchased items. Classify the extracted purchased items to identify the main categories of items purchased by users. Statistical analysis methods can be used to calculate the number and frequency of purchases for each item category. Record each purchased item and its related information to provide data support for subsequent purchase frequency calculation and demand analysis. Organize the information of each purchased item into a data table to ensure that the item information is associated with the user's purchase behavior for subsequent retrieval and analysis. Conduct a preliminary analysis of the purchase of items, identify the user's purchase preferences, and lay the foundation for subsequent personalized demand analysis. Select an appropriate frequency calculation method to determine the purchase frequency of each item. A simple counting method can be used to count the number of times each item appears in all purchase records. Determine the time range for frequency calculation (such as day, week, month) to ensure the comparability and consistency of the data. By grouping the extracted purchase items, count the number of purchases of each item. You can use Pandas' groupby method to group the data by item name, and then calculate the purchase frequency of each item. Record the purchase frequency of each item and generate a data table containing the item name, purchase frequency, and related statistical information. Choose a suitable demand analysis method, such as market basket analysis or personalized recommendation algorithm. These methods can help identify users' purchasing needs and preferences. You can use the association rule mining method to analyze the purchase relationship between different items and identify the potential demand for user purchases. Analyze users' purchasing needs based on item purchase frequency data. For example, you can calculate the probability of users purchasing a certain type of item and identify users' potential needs. By clustering the user's purchase behavior, identify the purchase demand characteristics of different user groups, and generate a personalized purchase demand model for users. Record the user's personalized purchase needs and related characteristics in the database to ensure the integrity and traceability of the data. At the same time, generate a demand analysis report to intuitively display the user's demand characteristics. Analyze the changing trend of user purchase needs to provide data support for subsequent user portrait fitting. Design a user portrait model, including multiple dimensions such as user basic information, purchasing habits, purchase needs, and preferences. Ensure that the model can fully reflect the user's consumption behavior and psychological characteristics. Determine the weight of each dimension so that the user's personalized characteristics can be accurately expressed during the fitting process. Fit the user portrait based on the user's multi-time point purchase habits and personalized purchase demand data. The weighted average method can be used to calculate the user's comprehensive score based on the weight of each dimension. Record each user's score in each dimension to form the final user portrait to ensure the accuracy and effectiveness of the portrait.The generated personalized user purchase portrait is saved in the database to ensure the integrity and searchability of the data. At the same time, visualization tools can be used to display the various dimensional features of the user portrait for easy understanding and analysis.

[0031] In this embodiment, step S4 includes the following steps: Step S41: locating the eye position of the detail-sharpened face image to obtain the user's eye position; Step S42: performing eye sight focus detection based on the user's eye position, thereby obtaining the user's sight focus; Step S43: performing time-series frame image segmentation on the detail-sharpened face image to obtain a continuous face image sequence; Step S44: tracking the gaze changes between consecutive frames of the continuous facial image sequence according to the user's gaze focus to extract the movement trajectory of the user's gaze focus.

[0032] In this embodiment, the detail-sharpened face image is preprocessed to improve the accuracy of eyeball positioning. Grayscale processing can be performed to reduce the computational complexity, and Gaussian blur is applied to remove noise to ensure the stability of subsequent processing. During the preprocessing process, the contrast and brightness of the image are adjusted to ensure that the eyeball area is clearly visible in the image, laying a good foundation for eyeball positioning. Select a suitable eyeball detection algorithm, such as Haar cascade classifier or Dlib's facial feature detection. The Haar cascade classifier is trained to locate the eye position with high accuracy, while Dlib's 68-point detection model can provide a more accurate eyeball position. According to the requirements of the algorithm, the image is adjusted as necessary, such as scaling or color conversion, to adapt to the input format of the algorithm. The processed detail-sharpened face image is analyzed using the selected eyeball detection algorithm to identify the position of the user's eyeball. The detection algorithm will return the coordinate frame of the eyeball area, usually composed of the coordinates of the upper left corner and the lower right corner. The identified eyeball position is recorded in the data to ensure that the eyeball position data of each user is associated with other facial features for subsequent processing. Choose a suitable gaze focus detection algorithm. Common ones include geometric models based on eye position or regression models based on machine learning. These algorithms can predict the user's gaze focus by analyzing the movement and position of the eye. Train the model with experimental data to ensure that it can work stably under different lighting and viewing angle conditions. Calculate the direction and focus of the gaze based on the user's eye position. Generally speaking, the determination of the gaze focus requires a combination of the angle of the eye and the direction of the head. You can use a three-dimensional geometric model to infer the user's gaze focus position. Record the coordinate information of the gaze focus and associate it with the eye position data to ensure that the user's focus point can be tracked. Analyze the obtained gaze focus to check its accuracy and stability in different scenarios. You can verify the effectiveness of the detection algorithm by setting a scene with a known focus. Save the gaze focus information in the database to provide data support for subsequent gaze change tracking. Get a real-time video stream from the camera and process each frame. Use image processing tools (such as OpenCV) to split the video stream into continuous frame images for time series analysis. Ensure the stability of the video stream and set an appropriate frame rate (such as 30 frames / second) to ensure that a sufficiently clear facial image sequence is captured. Save each frame in the video stream as an independent image file, and ensure the quality and clarity of each frame. A buffer can be set to temporarily store a certain number of frames for quick access in subsequent processing. Record the timestamp and sequence number of each frame to facilitate subsequent timing analysis and tracking. Save the segmented continuous face image sequence into a database, ensuring that each frame is associated with the timestamp and user information. Form a structured data set for subsequent analysis. Perform a preliminary analysis of the continuous face image sequence to identify the changes in the user's facial features at each time point, in preparation for subsequent tracking of gaze changes.Select a suitable gaze change tracking algorithm. You can use technologies such as optical flow or Kalman filter to analyze the gaze changes between consecutive frames. Optical flow can measure the movement of objects in the image, while Kalman filter is suitable for estimating the state of dynamic systems. According to the requirements of the algorithm, adjust the parameters to adapt to the user's gaze change characteristics to ensure the accuracy of tracking. In a continuous sequence of face images, use the selected tracking algorithm to analyze the user's gaze focus and extract the gaze movement trajectory. By comparing the gaze focus positions of adjacent frames, calculate the direction and speed of gaze movement. Record the gaze focus position at each time point and form a gaze movement trajectory diagram to show the user's gaze changes. Save the extracted gaze focus movement trajectory to the database to ensure the integrity and traceability of the data. At the same time, you can draw a gaze trajectory diagram to intuitively show the user's gaze movement. Analyze the gaze trajectory to identify the main areas and points of interest of the user, providing data support for subsequent user behavior analysis.

[0033] In this embodiment, step S5 includes the following steps: Step S51: performing a sight cross matching analysis on the user's sight focus movement trajectory based on the preset object area information to obtain a sight cross point; Step S52: Identify items at the intersection of the vending machine according to the intersection of the sight lines; Step S53: extracting the item information of the items at the intersection of the vending machine, and visualizing it in real time on the touch screen interface of the vending machine; Step S54: Calculate the user's gaze dwell time according to the touch screen interface of the vending machine to obtain the user's gaze dwell time.

[0034] In this embodiment, the preset area information of each item on the vending machine is determined. This information should include the coordinates, size and item name of each item on the touch screen of the vending machine. The area of ​​each item can be manually or automatically marked by using image processing software. A database containing all item area information is created for subsequent analysis. The accuracy of the area information is ensured to improve the accuracy of subsequent sight cross matching. Data is extracted from the user's sight focus movement trajectory to ensure that the timestamp and coordinate information of each sight focus position are complete. These data are sorted into a usable format, such as a time series data set, for cross matching analysis. The sight focus is sorted according to the timestamp to ensure that the movement path of the user's sight can be accurately tracked during the analysis process. The user's sight focus movement trajectory is cross-matched and analyzed with the item area information. A spatial matching algorithm can be used to determine whether the user's sight focus falls within a certain item area. A threshold (such as 5 pixels) is set to avoid erroneous matching caused by sight deviation. All sight focus positions that overlap with the item area are recorded, and sight intersection data is generated. These intersections will help identify the items of interest to the user on the vending machine. The data of the sight intersection is compared with the preset item area information to identify the specific items that the user's sight focuses on. Spatial query algorithms can be used to quickly locate the item area corresponding to each intersection. Record the name, category and other relevant information of the item at each intersection to form a data set containing the items that the user is paying attention to. Confirm the identified intersection items to ensure that they are within the time period when the user's line of sight appears. A time window (such as 1 second) can be set to ensure that the user's line of sight is indeed paying attention to the item during this time period. Count the frequency and duration of attention to each item for subsequent analysis of the user's purchase intention and preference. Record the information of the identified intersection items in the database to ensure the integrity and traceability of the data. At the same time, generate a report including user ID, line of sight intersection and paid attention to items for subsequent analysis. By analyzing the attention of intersection items, data support is provided for users' purchase decisions. According to the identified intersection items, relevant item information is extracted from the database, including item name, price, description and picture. This information will be used for subsequent visualization. Ensure that the extracted information is accurate and associated with the user's line of sight intersection so that it can be displayed on the touch screen. Use visualization tools to design the user interface of the vending machine touch screen to ensure that users can clearly see the information of the item when paying attention to it. The interface should include pictures, names, prices, and brief descriptions of items. Consider user experience when designing, and ensure that the interface is simple and easy to use, and that information is displayed intuitively. Graphical user interface (GUI) tools can be used for design and implementation. In the area of ​​the item that the user's line of sight focuses on, the touch screen interface is updated in real time to display the extracted item information. By setting a timer, item information is updated regularly to ensure that users see the latest information.Record the time of each information display and the corresponding item for subsequent analysis of user reactions and behaviors. Choose an appropriate method for calculating the gaze dwell time. You can use the timestamp difference calculation method, that is, calculate the dwell time by recording the time points when each gaze focus enters and leaves a certain item area. Set a threshold (such as 100 milliseconds) to ensure that only the time the user actually stays is recorded to avoid errors caused by short-term gaze movement. When the user's gaze enters a certain item area, record the current timestamp, and record the timestamp again when the gaze leaves the area. Calculate the time difference between the two, which is the user's gaze dwell time on the item. Count the gaze dwell time of each item, and record the average dwell time and total dwell time of the user on each item. Save the calculated gaze dwell time data in the database to ensure the integrity and traceability of the data. At the same time, generate a dwell time analysis report to show the user's attention to each item. By analyzing the user's gaze dwell time, provide important data support for subsequent user behavior analysis and personalized recommendations.

[0035] In this embodiment, step S6 includes the following steps: Step S61: predicting the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, thereby generating user instant purchase intention prediction data; Step S62: locating the vending machine item based on the user's instant purchase intention prediction data, and marking the vending machine item target; Step S63: performing sight error detection on the user's sight focus movement trajectory to obtain sight trajectory error data; Step S64: performing sight error correction on the vending machine item target based on the sight trajectory error data to obtain the final user tracking target.

[0036] In this embodiment, the user's sight dwell time is integrated with the personalized purchase portrait. This includes information such as the user's dwell time on each item, purchase frequency, historical purchase behavior, and user preferences. Ensure that the data format is unified for subsequent analysis. Preprocess the data, such as filling missing values ​​and removing outliers, to improve the quality and reliability of the data. The Z-score method or IQR method can be used to identify and remove outliers to ensure the accuracy of the data. Use machine learning algorithms (such as logistic regression, random forests, or support vector machines) to build a user purchase intention prediction model. Select appropriate features such as sight dwell time, purchase preferences in the user's personalized portrait, and historical purchase records as model inputs. Use historical data sets to train the model to ensure that the model can accurately identify the user's purchase intention. Through cross-validation and hyperparameter optimization, the generalization ability and accuracy of the model are improved. The latest user sight dwell time data is input into the trained model to generate the user's immediate purchase intention prediction data. These prediction data can quantify the user's purchase intention, such as expressing the possibility of a user purchasing a certain item in the form of probability. Record each user's immediate purchase intention data, including the intention probability and the corresponding item information, to provide a basis for subsequent item positioning and recommendation. According to the user's immediate purchase intention prediction data, set the standard for item target identification. For example, you can set an intention probability threshold (such as 70%), and only when the purchase intention probability of an item is higher than the threshold, it will be marked as a target item. By analyzing historical data, identify the user's attention characteristics for specific items to ensure the accuracy and reliability of target identification. Use the item area information of the vending machine and the user's immediate purchase intention prediction data to locate the items in the vending machine. You can use the coordinate matching method to compare the coordinates of the user's target item with the item area on the vending machine interface. Mark the items that meet the user's purchase intention to ensure that the user can see these items intuitively. The marking can be highlighted or highlighted with a border to attract the user's attention. Record the marked item target information in the database, including user ID, item name, positioning time, etc., for subsequent analysis and tracking. Evaluate the positioning results of the item target to analyze whether the user has shown interest in the marked item, and provide data support for subsequent user behavior analysis. Extract data from the user's line of sight focus movement trajectory to ensure that the timestamp and coordinate information of each trajectory point are complete. Organize these data into time series to facilitate subsequent error detection analysis. Confirm the integrity and consistency of the trajectory data to ensure that the results are not biased due to missing data during the analysis process. Select a suitable line of sight trajectory error detection algorithm. You can use a threshold-based method to set a line of sight deviation threshold (such as 10 pixels) and identify trajectory points that exceed the threshold. You can also consider using the dynamic time warping (DTW) algorithm to compare the difference between the user's line of sight trajectory and the ideal trajectory (such as the trajectory of the center point of the object) to identify errors in the trajectory.Record the identified sight trajectory errors, save the coordinates, timestamps and deviations from the ideal trajectory of the error points. These data will provide a basis for subsequent error correction. Analyze the error data and identify the main sources and patterns of errors, such as the frequency and direction of the user's sight deviation, so as to prepare for the subsequent correction steps. Design the corresponding error correction strategy based on the sight trajectory error data identified in the previous step. The weighted average method can be used to take the distance between the user's sight focus and the object target as the correction basis to gradually optimize the user's tracking target. Set correction parameters, such as the correction step size and the maximum number of corrections, to ensure the stability and effectiveness of the correction process. Perform sight error correction for each object target. By comparing the user's sight focus with the object target, calculate the actual sight deviation, and adjust the position or marking method of the object target according to the preset correction strategy. Record the results of each correction, including the sight focus position before and after correction, the error value and the adjusted object target information, to ensure the traceability of subsequent analysis. After completing the sight error correction, confirm the final user tracking target and record it in the database to ensure the integrity and traceability of the data. The recorded information includes user ID, corrected item target, and correction time, etc. Evaluate the final user tracking target, analyze the correction effect and user behavior response, and provide data support for subsequent behavior analysis and personalized recommendations.

[0037] In this embodiment, a user target tracking and analysis system based on AI visual recognition is provided, which is used to execute the user target tracking and analysis method based on AI visual recognition as described above, including: The image processing module is used to obtain real-time user purchase images based on the vending machine camera; perform multi-face positioning detection on the real-time user purchase images, and optimize the face details, thereby constructing a detail-sharpened face image; The identity intelligent perception module is used to obtain the historical monitoring logs of the vending machine; based on the historical monitoring logs of the vending machine, the user identity intelligent perception and user purchase behavior analysis are performed on the detail-sharpened face image to obtain optimized user purchase behavior data; Personalized portrait module, used to optimize user purchase behavior data, multi-time point purchase habit mining and user purchase demand analysis, so as to build a personalized purchase portrait of the user; The gaze tracking module is used to locate the eyeball position of the detail-sharpened face image and track the gaze changes between consecutive frames to extract the movement trajectory of the user's gaze focus; The sight cross matching module is used to perform sight cross matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculate the user's sight stay time, so as to obtain the user's sight stay time; The user intention prediction module is used to predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

[0038] The present invention helps to improve the clarity of the image through detail optimization technology, especially in the case of insufficient light or complex background. The optimized image ensures higher facial feature recognition accuracy and provides a clear basis for subsequent eye tracking and identity perception. In multi-person scenarios, the system can effectively identify and extract the facial features of each user, improve the accuracy and fault tolerance of the system, and work effectively even in dynamic environments. By combining historical behavior with real-time images, the system can intelligently identify users and track their historical behavior to achieve personalized recommendations. For example, the system can provide customized services for users based on their historical purchasing behavior, preferences, purchasing frequency and other information. Through the analysis of user purchasing behavior by deep learning and machine learning algorithms, it is possible to more accurately capture changes in user needs, thereby improving the relevance and accuracy of recommendations and optimizing the shopping experience. By analyzing purchasing behavior at different time points, it is possible to identify and predict the potential needs of users. For example, users may be more inclined to buy specific types of goods (such as holiday gifts, seasonal goods, etc.) in certain specific time periods, and the generation of personalized portraits helps to prepare for product recommendations in advance. By combining the user's purchasing habits, the vending machine can recommend products that are more in line with the user's interests and improve the conversion rate of purchase decisions. By accurately locating the eyeball position, the system can accurately track the user's sight focus and determine which products or areas the user is interested in. This provides basic data for subsequent product recommendations and user intention prediction. By continuously tracking the user's sight changes, the vending machine can adjust the product display in real time according to the user's attention changes, forming a more interactive and personalized shopping experience. By calculating the user's sight dwell time in the product area, the system can dynamically understand the user's purchasing interest and identify high-interest products. This helps the vending machine adjust the display content in real time, push products that the user is interested in, and improve the display effect of the product. Through the cross-matching analysis of sight, the system can not only identify the user's interest points, but also push information about related products or provide promotional reminders based on the dwell time to optimize purchase decisions. Combined with the sight dwell time and user portrait, the system can accurately predict the user's purchase intention and identify the user's purchase needs in advance. The timely prediction of user intention provides real-time data support for subsequent product recommendations and displays. By accurately predicting user intentions and accurately locating items, the vending machine can quickly respond to user needs and improve the purchase conversion rate. At the same time, the system can also optimize the product display order and the recommended information pushed in real time, thereby improving the user experience and the overall efficiency of the vending machine.

[0039] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0040] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A user target tracking and analysis method based on AI visual recognition, characterized in that: The following steps are involved: Step S1: Acquire real-time user purchase images based on the vending machine camera; Perform multi-face location detection on real-time user purchase images and optimize facial details to construct detailed sharpened facial images; Step S2: Obtain the historical monitoring log of the vending machine; perform intelligent user identity perception and user purchase behavior analysis on the detail-sharpened face image according to the historical monitoring log of the vending machine to obtain optimized user purchase behavior data; Step S3: Optimize the user's purchase behavior data to mine the purchase habits at multiple time points and analyze the user's purchase needs, so as to build a personalized purchase profile of the user; Step S4: locating the eyeball position of the detail-sharpened face image and tracking the sight line changes between consecutive frames to extract the movement trajectory of the user's sight line focus; Step S5: performing a sight cross-matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculating the user's sight stay time, thereby obtaining the user's sight stay time; Step S6: predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

2. The user target tracking and analysis method based on AI visual recognition according to claim 1 is characterized in that: The specific steps of step S1 are: Step S11: Acquire real-time user purchase images based on the vending machine camera; Step S12: performing multi-face location detection on the real-time user purchase image and marking multiple user face image frames; Step S13: calculating the face positions of multiple user face image frames one by one to obtain the distance of each face; Step S14: performing the closest distance recognition relative to the vending machine according to the distance of each face, and extracting the closest face image frame; Step S15: Optimizing the facial details of the nearest facial image frame, thereby constructing a detail-sharpened facial image.

3. The user target tracking and analysis method based on AI visual recognition according to claim 2 is characterized in that: The specific steps of step S15 are: Calculate the grayscale histogram of the nearest face image frame and extract the image grayscale histogram; Dynamically stretch the image grayscale histogram to obtain an adaptively stretched grayscale image; Perform image fusion reconstruction on the nearest face image frame according to the adaptively stretched grayscale image to construct a brightness optimized face image; Perform high-frequency noise detection on brightness-optimized face images and mark all image noise points; Calculate the pixel values ​​of all the image noise points; Performing adaptive filtering and noise reduction based on the pixel values ​​to obtain a filtered optimized face image; The face contour details of the filtered optimized face image are sharpened to construct a detail-sharpened face image.

4. The user target tracking and analysis method based on AI visual recognition according to claim 1 is characterized in that: The specific steps of step S2 are: Step S21: Obtaining the vending machine historical monitoring log; Step S22: performing facial geometric feature point recognition on the detail-sharpened facial image, and marking a plurality of facial geometric feature points; Step S23: Based on the historical monitoring log of the vending machine, intelligent user identity perception is performed on multiple facial geometric feature points to extract accurate user identity information; Step S24: Analyze the user's purchase behavior based on the user's accurate identity information and extract all the user's purchase behavior data; Step S25: Eliminate abnormal data from all user purchase behavior data to obtain optimized user purchase behavior data.

5. The user target tracking and analysis method based on AI visual recognition according to claim 1 is characterized in that: The specific steps of step S3 are: Step S31: performing purchase timestamp identification on the optimized user purchase behavior data to obtain the time point of each purchase behavior; Step S32: Evolving the periodic purchase behavior based on the time point of each purchase behavior to generate the user's periodic purchase pattern; Step S33: mining the multi-time point purchase habits of the user's periodic purchase rules, thereby obtaining the multi-time point purchase habits of the user; Step S34: Analyze the optimized user purchase behavior data to obtain the items purchased each time; Step S35: Calculate the purchase frequency of each item purchased, thereby generating the purchase frequency of each item; Step S36: Analyze the user's purchase demand according to the purchase frequency of each item to obtain the user's personalized purchase demand; Step S37: Perform user profile fitting based on the user's multi-time point purchasing habits and the user's personalized purchasing needs, thereby constructing a personalized purchasing profile of the user.

6. The user target tracking and analysis method based on AI visual recognition according to claim 1 is characterized in that: The specific steps of step S4 are: Step S41: locating the eye position of the detail-sharpened face image to obtain the user's eye position; Step S42: performing eye sight focus detection based on the user's eye position, thereby obtaining the user's sight focus; Step S43: performing time-series frame image segmentation on the detail-sharpened face image to obtain a continuous face image sequence; Step S44: tracking the gaze changes between consecutive frames of the continuous facial image sequence according to the user's gaze focus to extract the movement trajectory of the user's gaze focus.

7. The user target tracking and analysis method based on AI visual recognition according to claim 1 is characterized in that: The specific steps of step S5 are: Step S51: performing a sight cross matching analysis on the user's sight focus movement trajectory based on the preset object area information to obtain a sight cross point; Step S52: Identify items at the intersection of the vending machine according to the intersection of the sight lines; Step S53: extracting the item information of the items at the intersection of the vending machine, and visualizing it in real time on the touch screen interface of the vending machine; Step S54: Calculate the user's gaze dwell time according to the touch screen interface of the vending machine to obtain the user's gaze dwell time.

8. The user target tracking and analysis method based on AI visual recognition according to claim 1 is characterized in that: The specific steps of step S6 are: Step S61: predicting the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, thereby generating user instant purchase intention prediction data; Step S62: locating the vending machine item based on the user's instant purchase intention prediction data, and marking the vending machine item target; Step S63: performing sight error detection on the user's sight focus movement trajectory to obtain sight trajectory error data; Step S64: performing sight error correction on the vending machine item target based on the sight trajectory error data to obtain the final user tracking target.

9. A user target tracking and analysis system based on AI visual recognition, characterized in that: The method for performing the user target tracking and analysis method based on AI visual recognition as claimed in claim 1 comprises: The image processing module is used to obtain real-time user purchase images based on the vending machine camera; perform multi-face positioning detection on the real-time user purchase images, and optimize the face details, thereby constructing a detail-sharpened face image; The identity intelligent perception module is used to obtain the historical monitoring logs of the vending machine; based on the historical monitoring logs of the vending machine, the user identity intelligent perception and user purchase behavior analysis are performed on the detail-sharpened face image to obtain optimized user purchase behavior data; Personalized portrait module, used to optimize user purchase behavior data, multi-time point purchase habit mining and user purchase demand analysis, so as to build a personalized purchase portrait of the user; The gaze tracking module is used to locate the eyeball position of the detail-sharpened face image and track the gaze changes between consecutive frames to extract the movement trajectory of the user's gaze focus; The sight cross matching module is used to perform sight cross matching analysis on the movement trajectory of the user's sight focus based on the preset object area information, and calculate the user's sight stay time, so as to obtain the user's sight stay time; The user intention prediction module is used to predict the user's purchase intention based on the user's gaze dwell time and the user's personalized purchase portrait, and locate the items in the vending machine to obtain the final user tracking target.

Images