Chain store AI intelligent analysis and management method based on cloud-side cooperation

Abstract

The invention provides a chain store AI intelligent analysis and management method based on cloud edge collaboration. According to the method, data transmission bandwidth occupation can be effectively reduced, alarm response time is shortened to be within 100ms, collaborative optimization of local processing of sensitive data and a cloud model is realized, and real-time performance, safety and scene adaptability of chain store AI analysis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AIoT system integration and edge computing collaboration, and particularly relates to an AI intelligent analysis management method for chain stores based on cloud-edge collaboration. BACKGROUND

[0002] Although mainstream cloud-edge collaboration technology adopts a three-layer architecture of "cloud-end-edge-end" and realizes collaboration based on task splitting and data grading principles, it still has significant defects in the scene of Internet of Things and artificial intelligence fusion in chain stores and industries: at the architecture level, the cloud has huge pressure for global control and management, and it is difficult to cope with the dynamic access and unified management of a large number of heterogeneous edge nodes, the limited and heterogeneous resources of edge layer devices (AI super brain, gateway, intelligent IPC) lead to poor model deployment compatibility, and the explosive growth of terminal layer data acquisition devices causes data flooding; at the task and data collaboration level, the task splitting strategy is rigid, lacks real-time perception and dynamic adjustment capability for network status, device load and task priority, the data grading preprocessing mechanism is imperfect, the edge still uploads a large amount of redundant data, and cannot effectively alleviate the bandwidth pressure and cloud computing load; at the key technology implementation level, the model deployment and synchronization efficiency is low, the version management is chaotic, and the quality of feedback data is uneven, the task scheduling algorithm is not intelligent enough, the stability and security of lightweight communication protocols (MQTT, gRPC) in complex network environment are not good, the data consistency guarantee technology (incremental synchronization, breakpoint resume) has high implementation complexity and system overhead, and at the same time, the overall scheme still faces three core bottlenecks of high cost of massive data transmission, high delay of cloud processing at a long distance, and high security risk of sensitive data on the cloud, which cannot meet the business needs of real-time monitoring and immediate response. SUMMARY

[0003] The present application aims to at least solve one of the technical problems in the related art.

[0004] To this end, a first object of the present application is to provide an AI intelligent analysis management method for chain stores based on cloud-edge collaboration.

[0005] A second object of the present application is to provide an AI intelligent analysis management method for chain stores based on cloud-edge collaboration.

[0006] To achieve the above object, a first aspect of the present application provides an AI intelligent analysis management method for chain stores based on cloud-edge collaboration, comprising: S1, dynamically configuring a cloud deployment mode or an edge device deployment mode of an AI model according to real-time business needs and network status of chain stores; S2, based on the selected deployment mode, performing local real-time video stream analysis on the edge device, extracting feature information of target objects and generating structured analysis results; S3, uploading the structured analysis result to a cloud management platform through a lightweight communication protocol, while retaining the original video data locally on the edge device; S4, dynamically adjusting a model synchronization strategy and a task scheduling rule of the cloud and the edge according to the running state and the analysis result fed back by the edge device.

[0007] In an embodiment of the present application, the S1 further comprises: S11, making a deployment mode decision based on network state evaluation parameters and wherein represents a bandwidth occupation threshold value, represents a delay tolerance threshold value; S12, according to task priority classification and , assigning tasks to the edge device deployment by force, and assigning tasks to the cloud or the edge device according to a load balancing strategy.

[0008] In an embodiment of the present application, the S2 further comprises: S21, using a confidence filtering mechanism to screen the confidence of the detected target, and only retaining target objects; S22, performing a size filtering operation to constrain the size of the target frame, requiring pixels and pixels.

[0009] In an embodiment of the present application, the S3 further comprises: S31, using MQTT or gRPC protocol for structured data transmission, and the data packet size is not more than ; S32, using an incremental synchronization mechanism to upload only the model difference part, in combination with a breakpoint resume technology to ensure transmission integrity.

[0010] In an embodiment of the present application, the S4 further comprises: S41, according to the model running state fed back by the edge device, dynamically adjusting the model synchronization strategy , triggering model version rollback when the error rate is high; S42, reassigning the task priority of the cloud and the edge based on a task scheduling algorithm , and the algorithm input parameters include the current network state device load and task real-time requirement .

[0011] In an embodiment of the present application, further comprising: S5, performing a model verification step, uploading scene pictures not participating in training through a temporary verification mode , and based on the verification result adjusting the model parameters , wherein the verification result contains the confidence of the detection target and the annotation box coordinates .

[0012] To achieve the above purpose, the second aspect of the present application provides a cloud edge collaborative based AI intelligent analysis management device for chain stores, comprising: A deployment mode configuration module is configured to dynamically configure the cloud deployment mode or the edge device deployment mode of the AI model according to the real-time business requirements and network status of the chain stores. A video stream analysis module is configured to perform local real-time video stream analysis on the edge device based on the selected deployment mode, extract feature information of the target object, and generate structured analysis results. A data transmission module is configured to upload the structured analysis results to the cloud management platform through a lightweight communication protocol, while retaining the original video data locally on the edge device. A task scheduling adjustment module is configured to dynamically adjust the model synchronization strategy and task scheduling rules of the cloud and the edge based on the running state and analysis results fed back by the edge device.

[0013] The method and device of the present application effectively reduce data transmission bandwidth occupation and shorten response delay, realize local processing of sensitive data and cloud aggregation of intelligent analysis results through a cloud edge collaborative architecture, and improve the real-time performance and security of AI analysis for chain stores.

[0014] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart of a cloud edge collaborative based AI intelligent analysis management method for chain stores is provided for the embodiments of the present application; Figure 2 An AI system architecture diagram is provided for the embodiments of the present application; ​Figure 3 An AI deployment scheme architecture diagram provided for an embodiment of the present application; Figure 4 A specific flowchart of an AI application provided for an embodiment of the present application; Figure 5 A vehicle network diagram against external damage provided for an embodiment of the present application, and a comparison diagram of actual point positions photographed by a mobile phone; Figure 6 A pixel requirement diagram provided for an embodiment of the present application; Figure 7 A target dense or large inclination angle suitable for quadrilateral labeling diagram provided for an embodiment of the present application; Figure 8 A diagram for accurately defining a target to be recognized provided for an embodiment of the present application; Figure 9 A first labeling target diagram provided for an embodiment of the present application; Figure 10 A second labeling target diagram provided for an embodiment of the present application; Figure 11 A third labeling target diagram provided for an embodiment of the present application; Figure 12 A fourth labeling target diagram provided for an embodiment of the present application; Figure 13 Three deployment mode architecture diagrams provided for an embodiment of the present application; Figure 14 An edge deployment scheme topology diagram provided for an embodiment of the present application; Figure 15 A cloud deployment scheme topology diagram provided for an embodiment of the present application; Figure 16 A center deployment scheme topology diagram provided for an embodiment of the present application; Figure 17 A rule design diagram of an AI application provided for an embodiment of the present application; Figure 18 An algorithm arrangement actual diagram of an AI application provided for an embodiment of the present application; Figure 19 A target relationship processing configuration diagram provided for an embodiment of the present application; Figure 20 A red light on condition judgment diagram provided for an embodiment of the present application; Figure 21 A structure diagram of a chain store AI intelligent analysis management device based on cloud edge collaboration provided for an embodiment of the present application. DETAILED DESCRIPTION

[0016] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict. The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0017] In order for those skilled in the art to better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0018] A cloud-edge collaborative-based chain store AI intelligent analysis management method and device are described below with reference to the drawings according to an embodiment of the present application.

[0019] Embodiment 1 Figure 1 The flowchart of the cloud-edge collaborative-based chain store AI intelligent analysis management method according to the embodiment of the present application is shown as Figure 1 , which includes: S1, dynamically configuring a cloud deployment mode or an edge device deployment mode of an AI model according to real-time business requirements and network states of chain stores.

[0020] Specifically, in some implementations, dynamically configuring a cloud deployment mode or an edge device deployment mode of an AI model according to real-time business requirements and network states of chain stores is a key scheduling mechanism in the cloud-edge collaborative architecture of the present application. This step dynamically switches the deployment location of the AI model between the cloud and the edge through an intelligent decision engine, in combination with the current business characteristics (such as customer flow density, task type, response time requirement) and network states (such as bandwidth, delay, packet loss rate) of the store, to achieve optimal resource utilization and business response efficiency.

[0021] This mechanism relies on a real-time monitoring module and a task scheduling strategy engine. The monitoring module continuously collects CPU / GPU load, network bandwidth usage, task response delay, and other indicators of the edge device, and reports the current business requirements of the store (such as whether it is in a peak period, whether it needs real-time alarm) to the cloud scheduling center. The scheduling strategy engine dynamically selects the deployment mode based on a preset decision model (such as a rule-based decision tree or a machine learning-based prediction model). For example, when the store is in a high customer flow period and the edge device load is below a threshold , the system deploys the AI model to the edge device to achieve millisecond-level response; when the network bandwidth is sufficient and the task has a high non-real-time requirement, the system schedules the task to the cloud for execution to utilize stronger computing resources.

[0022] Multiple key parameters are involved in the scheduling strategy, including but not limited to: minimum response time of edge device (usually within 100 ms) , minimum bandwidth requirement of cloud deployment (suggested ≥ 10 Mbps), task priority (such as alarm task priority is 1, statistical task is 3), upper limit of model inference delay (such as video event analysis model delay should be ≤ 500 ms), etc. In addition, the system supports configuring the minimum confidence threshold of edge deployment (such as ≥ 80%), to ensure the reliability of edge inference results.

[0023] This mechanism is widely applicable to scenarios such as back-of-house compliance detection, security monitoring, and personnel identification in chain stores. For example, in the back-of-house scenario, when it is detected that an employee is not wearing a mask and the edge device load is low, the system can deploy the model to the edge to achieve instant alarm; during off-peak hours, the system can upload data to the cloud for batch analysis to optimize model performance.

[0024] This step effectively solves the high delay and high bandwidth consumption problems caused by traditional centralized cloud processing, while avoiding the response delay caused by insufficient computing power at the edge. Through dynamic deployment strategy, the system ensures real-time performance while significantly reducing data transmission costs and improving the robustness and adaptability of overall AI analysis.

[0025] Further, S1 includes: S11, making deployment mode decision based on network state evaluation parameters and , wherein represents a bandwidth occupancy threshold, represents a delay tolerance threshold.

[0026] Specifically, in some implementations, deployment mode decision is made based on network state evaluation parameters and . This step quantifies the boundary conditions of network bandwidth occupancy and delay tolerance, enabling intelligent switching between cloud analysis and edge device local analysis mode, thereby optimizing system resource utilization efficiency while ensuring real-time performance.

[0027] This step first collects the current network bandwidth occupancy rate and task response delay in real time through the network state monitoring module. Bandwidth occupancy rate is usually calculated as the ratio of data uploaded / downloaded per unit time (such as Mbps) to the maximum bandwidth capacity of the network, and delay The evaluation is based on the average time (in milliseconds) from task triggering to result return. With the preset bandwidth usage threshold If a comparison is made, If this is detected, the system determines that the current network bandwidth is under high load and should prioritize edge analysis mode to reduce data transmission pressure. Simultaneously, the system will... With delay tolerance threshold If a comparison is made, If the current network latency is within an acceptable range, the cloud-based analysis mode can continue to be used to obtain more accurate recognition results.

[0028] It is typically set to 60% - 80% of the network bandwidth capacity to ensure that basic communication capabilities can still be maintained in high-concurrency scenarios; This is set according to business needs. For example, in security alarm scenarios, the latency tolerance is usually no more than [missing information]. However, in non-real-time statistical tasks, the restrictions can be relaxed to... The above parameters can be dynamically adjusted through the system configuration interface to adapt to the network environment and business priorities of different stores.

[0029] This step is widely used in scenarios such as compliance inspections in the kitchens of chain stores, personnel identification, and area intrusion alarms. For example, in a kitchen setting, if network bandwidth utilization exceeds... The system will automatically switch to edge devices for real-time identification and alarms, uploading only structured data to the cloud, thereby significantly reducing bandwidth consumption and improving response speed. When network conditions are good, the system can enable cloud analysis mode to obtain more accurate identification results, such as multi-attribute classification or high-precision video event analysis.

[0030] This step, by introducing a quantitative evaluation mechanism, enables dynamic adaptation of deployment modes, effectively mitigating the performance degradation caused by network fluctuations in traditional cloud-edge collaborative systems. Its innovation lies in combining network state parameters with business requirements to construct a system based on… and The intelligent decision-making model improves overall analysis efficiency and resource utilization while ensuring system stability.

[0031] S12, Classified by task priority and ,Will The task will be forcibly assigned to edge device deployment. Tasks are distributed to cloud or edge devices according to load balancing strategies.

[0032] Specifically, in some implementations, tasks are categorized according to priority. and ,Will Tasks are forcibly assigned to edge device deployments, and... Tasks are allocated to cloud or edge devices according to a load balancing strategy, which is the key task scheduling mechanism in the cloud-edge collaborative architecture of this invention. This mechanism combines task priority allocation with dynamic resource scheduling to achieve efficient, real-time, and secure processing of AI intelligent analysis tasks for chain stores.

[0033] Task Priority Classification and This is a quantitative assessment based on dimensions such as task response time requirements, data sensitivity, and processing complexity. For example, tasks with high real-time requirements and those involving privacy-sensitive data (such as store customer flow identification and personnel behavior monitoring) are categorized as... These types of tasks must be executed on edge devices to ensure that data does not leave the local machine, meeting compliance and low-latency requirements. The tasks include non-real-time tasks (such as historical data statistics and trend analysis), whose processing results have lower real-time requirements and can be dynamically distributed to the cloud or edge devices based on the current load status of the edge devices using a load balancing strategy.

[0034] Task priority classification needs to be based on specific business scenarios and threshold settings. For example, the response time threshold can be set as follows: If the task's estimated response time Then it is classified as ;like Then it is classified as Furthermore, the load status of edge devices can be monitored via CPU utilization. Memory usage Task queue length The cloud scheduler evaluates tasks based on metrics such as weighted round-robin or minimum load priority, and then allocates tasks accordingly.

[0035] This mechanism is widely applicable to scenarios such as kitchen compliance management, security monitoring, and personnel identification in chain stores. For example, in the kitchen scenario, the task of identifying employees not wearing masks falls under... The task of processing and triggering alarms in real time on edge devices is required; while the task of counting daily customer traffic in stores falls under... It can be centrally processed by the cloud during off-peak hours to reduce the resource consumption of edge devices.

[0036] This step effectively solves the problems of high latency, high bandwidth consumption, and data security risks in traditional centralized cloud processing architectures. Through mandatory... Tasks are executed at the edge, ensuring real-time response and localized data processing; while Load balancing of tasks improves the overall system's resource utilization and task processing efficiency, thereby achieving comprehensive optimization of the intelligent analysis system under cloud-edge collaboration in terms of performance, cost, and security.

[0037] S2 performs local real-time video stream analysis on edge devices based on the selected deployment mode, extracts feature information of target objects, and generates structured analysis results.

[0038] Specifically, in some implementations, based on the selected deployment mode, local real-time video stream analysis is performed on edge devices to extract feature information of target objects and generate structured analysis results. This step achieves low-latency processing and feature extraction of video streams by deploying AI models at the edge, thereby completing intelligent analysis tasks without relying on cloud computing, significantly reducing data transmission bandwidth consumption and improving response efficiency.

[0039] This step first relies on the hardware capabilities of the edge device, typically an AI box or smart IPC with GPU acceleration. The edge device accesses the real-time video stream via standard protocols such as RTSP and ONVIF, performing video decoding and processing at a set frame rate (e.g., 15fps or 30fps). During video stream processing, the AI ​​model infers from each frame, identifying the target object (e.g., people, vehicles, objects) and extracting its key feature information, such as location coordinates, category labels, confidence levels, and attribute values ​​(e.g., color, state). The inference results are output in a structured data format (e.g., JSON), containing metadata such as the target object's identification information, timestamp, and spatial location.

[0040] Video analytics on edge devices must meet certain pixel requirements. For example, for object detection models, the bounding box should be no smaller than 48×48 pixels at 2 megapixel resolution and no smaller than 64×64 pixels at 4 megapixel resolution. Furthermore, the confidence threshold during inference is typically configured to be above 80% to ensure the reliability of the recognition results. Simultaneously, edge devices support various bounding box types (such as rectangles and quadrilaterals) to adapt to the recognition needs of targets with different shapes. Regarding detection region configuration, users can define rectangular or custom polygonal regions based on the actual scene to limit the analysis range of the AI ​​model, improving analysis accuracy and reducing false positives.

[0041] At the application level, this step is widely used in scenarios such as kitchen compliance management, security monitoring, and personnel behavior recognition in chain stores. For example, in the kitchen scenario, edge devices can detect in real time whether employees are wearing masks, hats, or uniforms, identify abnormal behavior, and generate structured event records. In the security scenario, edge devices can instantly identify and alarm on events such as intrusion and exceeding personnel limits, without waiting for a cloud response, thus achieving millisecond-level local decision-making.

[0042] The technical benefits of this step are reflected in several aspects: First, by performing real-time analysis locally, the latency of cloud processing is significantly reduced, meeting the needs of high real-time business operations; second, by uploading only structured results instead of the original video stream, network bandwidth resources are effectively saved; third, sensitive data is processed at the edge, enhancing data privacy and security; and finally, the structured analysis results facilitate subsequent cloud aggregation, statistics, and model iteration optimization, forming a closed loop to improve the overall performance of the system.

[0043] Furthermore, S2 includes: S21 employs a confidence filtering mechanism to assess the confidence level of the detected target. Filter and keep only The target object.

[0044] Specifically, in some implementations, a confidence filtering mechanism is used to assess the confidence level of the detected target. Filtering is a crucial step in the Intelligent Event Detection (IED) process. Its technical principle is based on the confidence value output by the target detection model, used to filter out false positives with low confidence, thereby improving the accuracy of detection results and system processing efficiency. This step is typically implemented in an AI inference engine on edge devices or in the cloud, working in conjunction with parameters such as detection area and triggering rules as part of the intelligent analysis task configuration.

[0045] This step involves setting a confidence threshold. The model evaluates each object detection result output by it. Specifically, after completing inference for an image or video frame, the model outputs a confidence value for each detected object. This represents the probability that the target belongs to a certain category. The system will iterate through all detection results and retain only those that satisfy the criteria. The target object. This threshold can be dynamically adjusted according to the actual application scenario. For example, in security scenarios with high precision requirements, the threshold can be increased to... However, under complex backgrounds or low-light conditions, the setting can be appropriately reduced to [specific value]. To improve the recall rate.

[0046] Confidence It is usually expressed as a percentage, with a range of... In this invention, the minimum threshold for the confidence filtering mechanism is set to 30%, a value determined based on the balance between actual test data and the false positive rate (FPR). In actual deployment, this threshold can be set through the IED rule configuration interface, allowing users to flexibly adjust it according to business needs. For example, in "Regional Target Anomaly Detection," users can configure... This ensures that only anomalous targets with high confidence are reported.

[0047] This step is widely used in AI-powered intelligent analysis systems for chain stores, such as kitchen compliance checks, personnel identification, and area intrusion alarms. Taking "personnel entering the work area without a safety helmet" as an example, after detecting the target, the system first filters out individuals using a confidence-based filtering mechanism. The system targets specific objectives and combines this with area detection and duration assessment to achieve accurate alarms. This mechanism effectively reduces the processing burden of false alarms and redundant data, especially when deployed at the edge, helping to reduce the consumption of local computing resources and improve real-time response capabilities. This step significantly improves the robustness and accuracy of the AI ​​analysis system. By filtering low-confidence targets, the system can reduce the triggering of invalid events caused by model misidentification, thereby improving the credibility and processing efficiency of alarms. In addition, this mechanism also supports linkage with parameters such as "conditional duration sensitivity" to achieve more complex logical judgments, such as triggering an alarm only when the target confidence level remains above a threshold for a set time, further enhancing the system's ability to identify real events.

[0048] S22, Perform size filtering operation on the target frame size. To impose constraints, require pixels and Pixel.

[0049] Specifically, in some implementations, performing size filtering is a key step in configuring Intelligent Event Detection (IED) rules. Its technical principle is based on the bounding box size information output by the object detection model, filtering identified targets to eliminate those that might be misidentified or have reduced recognition accuracy due to insufficient image resolution or targets that are too small. Specifically, this step involves setting the minimum width of the target bounding box. and minimum height ,Require pixels and Pixels, thus ensuring that the target has enough pixel information in the image for subsequent identification and analysis.

[0050] This size filtering operation is typically embedded in the IED rule engine as preprocessing logic after target recognition. On edge devices or cloud servers, the bounding box information (i.e., the target's coordinates and dimensions) output by the target detection model is read and evaluated in real time. If the detected target's width or height is below a set threshold, the target is filtered out and not included in subsequent rule evaluations or event triggering. This operation is usually performed in parallel within the video frame processing flow to minimize the impact on overall processing performance.

[0051] The threshold setting for size filtering needs to be adjusted based on the resolution of the image acquisition device. For example, at a resolution of 2 megapixels, the target bounding box size needs to meet certain requirements. Pixels, and at 4 megapixel resolution, this threshold increases to Pixels. This setting meets the basic recognition requirements of deep learning models for target size, ensuring sufficient robustness and recognition accuracy in practical deployment.

[0052] This step is widely used in scenarios such as compliance inspections in the kitchens of chain stores, personnel identification, and security monitoring. For example, when identifying personnel not wearing safety helmets, if the target is too small (e.g., less than 48 pixels), the model may not be able to accurately identify its attributes, leading to false positives or false negatives. Size filtering can effectively improve the accuracy of identification and the stability of the system.

[0053] Furthermore, this step has significant technical value in cloud-edge collaborative architectures. By constraining the target size at the edge or in the cloud, the processing overhead of invalid targets can be reduced, improving the overall system response efficiency and resource utilization. At the same time, it can reduce the false alarm rate caused by the instability of small target recognition, thereby enhancing the practicality and deployment feasibility of AI intelligent analysis systems.

[0054] S3 uploads the structured analysis results to the cloud management platform via a lightweight communication protocol, while retaining the original video data locally on the edge device.

[0055] Specifically, in some implementations, the structured analysis results are uploaded to the cloud management platform through a lightweight communication protocol, while the original video data is retained locally on the edge device. This is the key data transmission and storage mechanism in the cloud-edge collaborative architecture of this invention, which aims to achieve an efficient, secure, and low-latency intelligent analysis closed loop.

[0056] This step employs lightweight communication protocols (such as MQTT and gRPC) to achieve efficient data interaction between edge devices and the cloud platform. After completing local AI analysis, the edge device only encapsulates and uploads structured data (such as event type, timestamp, confidence level, target location, etc.) to the cloud, while the raw video data is stored locally at the edge, typically using a circular buffer mechanism or a time / event-triggered local storage strategy. The structured data upload process supports breakpoint resumption and incremental synchronization mechanisms to ensure data transmission can be resumed after network instability or interruption, while avoiding duplicate uploads and improving transmission efficiency.

[0057] The upload frequency of structured data can be configured according to the task type. For example, in edge device local analysis mode, the minimum upload interval can be set to 1 minute; while in cloud analysis mode, the minimum interval is 5 minutes. The uploaded data is usually in JSON or Protobuf format, containing key fields such as event type (e.g., "not wearing a safety helmet"), number of targets, duration (in seconds), and confidence level (range 30% - 100%). Video data stored locally on edge devices is usually compressed and stored in H.264 or H.265 encoding format, and the storage period can be configured from 24 hours to 7 days, depending on the device's storage capacity and business needs.

[0058] This process is widely used in scenarios such as kitchen compliance management, personnel identification, and security monitoring in chain stores. For example, in a kitchen setting, edge devices detect in real time whether employees are wearing masks, hats, or uniforms, uploading only structured information about abnormal events to the cloud, while the original video is stored locally for subsequent manual review or local backtracking. In a security setting, edge devices identify intrusion behavior or personnel exceeding limits, uploading the event type and timestamp to the cloud while retaining complete video recordings to meet compliance audit and event tracing requirements.

[0059] By uploading only the structured analysis results instead of the raw video stream, this step significantly reduces network bandwidth usage, improves data transmission efficiency, and reduces cloud storage pressure. The edge device retains the original video data locally, ensuring data integrity and traceability, making it particularly suitable for scenarios with high requirements for data privacy and compliance.

[0060] Furthermore, S3 includes: S31, use MQTT or gRPC protocol for structured data transmission, with packet size not exceeding [a certain value]. .

[0061] Specifically, in some implementations, this invention employs MQTT or gRPC protocols for structured data transmission to ensure efficient and low-latency data interaction in a cloud-edge collaborative architecture. This step is technically implemented based on the characteristics of IoT communication protocols, combined with the data transmission requirements between edge devices and the cloud. It utilizes the lightweight, low-power, and highly reliable MQTT protocol, or the high-performance, low-latency gRPC protocol that supports streaming, to achieve structured data exchange between the edge and the cloud.

[0062] The data packet size is strictly limited to no more than This approach adapts to the limited bandwidth and storage capabilities of edge devices. This limitation ensures that after local AI analysis at the edge, only structured results (such as event tags, target locations, confidence levels, etc.) need to be uploaded, rather than raw video streams or high-resolution images, thus significantly reducing network load. The MQTT protocol supports QoS (Quality of Service) level configuration; typically, QoS 1 or QoS 2 is used in this solution to ensure the reliability and orderliness of data transmission. The gRPC protocol, through a binary transmission mechanism based on HTTP / 2, achieves more efficient serialization and deserialization operations, suitable for scenarios requiring real-time feedback.

[0063] This step is widely used in AI-powered intelligent analysis systems for chain stores, such as kitchen compliance inspection, personnel identification, and security monitoring. Edge devices (such as AI boxes and smart IPCs) complete AI inference tasks locally, including object detection, classification, and OCR, and then encapsulate the structured data into JSON or Protobuf format, uploading it to the cloud management platform via MQTT or gRPC protocols. Upon receiving the data, the cloud platform performs event aggregation, statistical analysis, and model iteration and optimization, then distributes the updated model back to the edge devices, forming a closed-loop optimization mechanism.

[0064] This step, through a structured data transmission mechanism, effectively solves the high bandwidth consumption problem caused by traditional video streaming, while ensuring data consistency and real-time performance between the edge and the cloud. (Data packet size limit) Combined with protocol selection, the system can still operate stably in low-bandwidth environments, improving the overall system's deployment flexibility and response efficiency, and providing reliable data transmission guarantees for the intelligent operation of chain stores.

[0065] S32, through incremental synchronization mechanism Upload only the differences in the model, combined with breakpoint resume technology. Ensure transmission integrity.

[0066] Specifically, in some implementations, the present invention uses an incremental synchronization mechanism. With breakpoint resume technology The combination of these technologies enables efficient uploading and transmission integrity assurance of model differences. This step plays a crucial role in the cloud-edge collaborative architecture, especially when deploying model updates on edge devices, effectively reducing bandwidth consumption and improving synchronization efficiency.

[0067] Incremental synchronization mechanism By comparing the differences between the models currently deployed on edge devices and the latest models delivered from the cloud, only the differing parts (such as weight updates and structural changes) are uploaded, rather than the entire model file. This mechanism is typically based on version control and hash verification technology, determining the model changes that need to be synchronized by comparing the model version number with the hash values ​​of key parameters. Resumable upload technology is also included. During model upload, if transmission fails due to network interruption, device disconnection, or other reasons, the system can record the current transmission progress and resume uploading from the point of interruption after the connection is restored, avoiding duplicate transmissions and resource waste. This technology relies on the maintenance of chunked transmission protocols (such as HTTP Range requests) and local transmission status logs.

[0068] In incremental synchronization, model version numbers typically employ semantic versioning (e.g., v1.2.3), and difference calculations are based on binary differential algorithms for model files (e.g., rsync, xdelta). Resumable uploads support a maximum block size of 1MB to 10MB, with transmission status logs recorded at 100KB granularity, ensuring accurate recovery after interruption. Furthermore, edge devices support a maximum of 5 concurrent upload tasks, and the average transmission latency for a single model synchronization task is controlled within 500ms, making it suitable for edge AI scenarios with high real-time requirements.

[0069] This step is widely used in AI model deployment and update scenarios in chain stores, such as detection of kitchen staff wearing protective gear, store traffic statistics, and security intrusion detection. In edge devices (such as AI supercomputers and intelligent IPCs), model updates are typically issued through industry application platforms. Upon receiving the update command, the edge device executes the update. Difference calculation, and through The mechanism ensures reliable uploading of model files. This mechanism is particularly suitable for edge deployment environments with unstable networks and limited bandwidth, such as remote stores or mobile edge nodes.

[0070] pass and Through collaboration, this invention significantly reduces the amount of data transmission required for model synchronization, improves synchronization efficiency, and ensures the integrity and reliability of model updates. In actual deployment, this mechanism can reduce model update bandwidth consumption by more than 60% and ensure a success rate of no less than 99.5% for model updates in weak network environments, thereby improving the overall stability and response speed of the system.

[0071] S4 dynamically adjusts the model synchronization strategy and task scheduling rules between the cloud and the edge based on the operating status and analysis results fed back by the edge devices.

[0072] Specifically, in some implementations, the model synchronization strategy and task scheduling rules between the cloud and the edge are dynamically adjusted based on the operating status and analysis results fed back by the edge devices. This step achieves intelligent collaboration and dynamic optimization between the cloud and the edge by sensing the resource load, network status, task execution efficiency, and AI model inference results of the edge devices in real time, thereby improving the overall system's response speed, resource utilization, and model iteration efficiency.

[0073] This step relies on a two-way communication mechanism between the edge device and the cloud. When performing local AI inference tasks, the edge device periodically uploads its running status (such as CPU / GPU utilization, memory usage, and task queue length) and inference results (such as identification confidence, number of targets, and frequency of abnormal events) to the cloud management platform via lightweight protocols (such as MQTT and gRPC). Based on the received feedback information and pre-defined scheduling strategies (such as QoS-based priority scheduling and load-balanced resource allocation), the cloud management platform dynamically adjusts the model synchronization strategy and task scheduling rules. For example, when the edge device reports that its GPU load exceeds a threshold of 80% and its inference latency exceeds a certain threshold... At the same time, some non-real-time tasks can be migrated to the cloud for execution, while model updates can be paused or delayed to avoid further overload of edge devices due to frequent model loading.

[0074] This step involves configuring and evaluating several key parameters. The feedback frequency of the edge device is typically set to... to This is done once to ensure a balance between the real-time nature of status information and system overhead. In the cloud-based scheduling algorithm, task priority is determined based on task type (e.g., alarm, statistics, identification) and business requirements (e.g., real-time performance, accuracy), combined with the remaining computing power of edge devices. With network bandwidth Dynamic decision-making is performed. In the model synchronization strategy, the cloud will make decisions based on the model version differences of the edge devices. Compared with model update revenue The decision on whether to trigger model distribution, among which The model validation results can be compared with the historical false alarm rate of edge devices via cloud-based models. Conduct a quantitative assessment.

[0075] This process is widely used in scenarios such as compliance inspections in the kitchens of chain stores, personnel identification, and security monitoring. For example, in a kitchen setting, edge devices are responsible for real-time detection of whether employees are wearing masks, hats, etc., while the cloud platform uses the accuracy of the edge devices' identification to determine the correct settings. With false alarm rate Dynamically adjust the model synchronization strategy to ensure continuous optimization of the model at the edge, while avoiding business interruptions caused by frequent model updates.

[0076] The technical benefits of this step are a significant improvement in the adaptive capabilities and resource utilization efficiency of the cloud-edge collaborative system. By dynamically adjusting the model synchronization strategy, the computing and storage pressure on edge devices can be effectively reduced, extending the stability of device operation. Through the optimization of task scheduling rules, task load balancing and response priority control can be achieved, thereby improving the overall system's intelligence level and business closed-loop capabilities.

[0077] Furthermore, S4 includes: S41, based on the model's operating status fed back by the edge device. Dynamically adjust the model synchronization strategy ,when medium error rate The model version rollback is triggered at that time.

[0078] Specifically, in some implementations, the model's operating status is based on feedback from edge devices. Dynamically adjust the model synchronization strategy This is the key mechanism for achieving closed-loop optimization of the cloud-edge collaborative AI intelligent analysis and management system in this invention. This step involves real-time monitoring of the model execution status of edge devices, combined with a preset error rate threshold. This triggers a model version rollback operation, thereby ensuring the stable operation and accurate inference of the AI ​​model at the edge.

[0079] After deploying AI models, edge devices will continuously report their operational status. Including inference delay, recognition accuracy, and error rate. Key indicators such as resource utilization are monitored. The cloud management platform receives and parses this status information through lightweight communication protocols (such as MQTT and gRPC), and combines this information with the hardware performance, network conditions, and task priorities of edge devices to dynamically decide on synchronization strategies. This includes whether to update the model version, switch deployment modes (cloud analytics / edge analytics), and perform model rollback. When detected... When this happens, the system will trigger the model version rollback mechanism to revert the current model version on the edge device to the previous stable version in order to quickly restore the accuracy and stability of model inference.

[0080] Error rate The calculation is based on the ratio of the number of recognition errors by the edge device within a unit time window (e.g., 10 minutes) to the total number of recognitions, i.e. ,in This indicates the number of incorrectly identified targets. This indicates the total number of targets identified. (System settings) This threshold, which serves as the rollback trigger threshold, can be configured according to different business scenarios. For example, in security scenarios with high precision requirements, it can be set to [value missing]. However, in general monitoring scenarios, the restrictions can be appropriately relaxed to... .

[0081] This mechanism is widely applicable to AI analysis tasks in chain stores, such as compliance inspections in the kitchen, personnel identification, and security monitoring. For example, in a kitchen setting, if the error rate of a personnel wearable detection model deployed on edge devices increases after a version update... The system will automatically roll back to the previous version to ensure accurate identification of key behaviors such as whether employees are wearing masks and hats, and avoid false alarms or missed alarms due to degraded model performance.

[0082] This step enables adaptive optimization and fault-tolerant recovery of the AI ​​model at the edge, effectively improving the system's robustness and availability. Through dynamic adjustment... The system can respond quickly when model performance degrades, reducing business interruptions caused by model failure, while also reducing the frequency of manual intervention and improving operational efficiency. Furthermore, this mechanism supports integration with the model training platform, sending error samples from the edge back to the cloud for subsequent model iteration and optimization, forming a closed-loop process of "deployment-running-feedback-optimization," significantly enhancing the AI ​​model's adaptability and continuous evolution capabilities in complex edge environments.

[0083] S42, based on task scheduling algorithm The algorithm reprioritizes tasks between the cloud and the edge, with input parameters including the current network state. Equipment load and the real-time requirements of tasks .

[0084] Specifically, in some implementations, based on task scheduling algorithms The algorithm reallocates task priorities between the cloud and the edge. It is designed to adjust priorities based on the current network status. Equipment load and the real-time requirements of the task It dynamically adjusts the execution priority of tasks in the cloud and at the edge, thereby optimizing system resource utilization and network bandwidth consumption while ensuring task response efficiency.

[0085] A multi-dimensional evaluation mechanism is employed to make real-time decisions on the execution strategy for each task. Specifically, the algorithm first collects the current network state. This includes metrics such as bandwidth utilization, latency, and packet loss rate to assess the communication capabilities between the edge and the cloud. Secondly, device load... This includes metrics such as CPU utilization, GPU utilization, and memory usage of edge devices, used to determine whether the edge has sufficient computing power to execute the current task. Finally, there are real-time requirements for the task. Depending on the type of task, such as video event analysis or intrusion detection, the required response time is typically in the range of milliseconds to seconds.

[0086] Furthermore, Tasks are prioritized using a weighted scoring model. For example, for tasks with high real-time requirements, if the edge device has low load and good network conditions, it is executed first at the edge to reduce cloud transmission latency. Conversely, if the edge device has high load or poor network conditions, the task is scheduled to be executed in the cloud to ensure task completion quality. Optionally, the algorithm supports dynamically adjusting the weight coefficients to adapt to the prioritization strategy under different business scenarios.

[0087] This step is widely used in intelligent monitoring systems for chain stores, such as kitchen compliance checks, personnel identification, and security alarms. By adjusting the task execution location in real time, the system can reduce cloud computing pressure and network transmission costs while ensuring business response speed, thereby improving the overall system's robustness and scalability.

[0088] Also includes: S5, which performs the model validation step via temporary validation mode. Upload scene images that were not used in the training. And based on the verification results Adjusting model parameters The verification result includes the confidence level of the detected target. and coordinates of the annotation box .

[0089] Specifically, in some implementations, performing model validation is a crucial step in model optimization and performance evaluation within AI intelligent analysis systems. This step utilizes a temporary validation mode. Upload scene images that were not used in the training. And based on the verification results Adjusting model parameters This improves the detection accuracy and robustness of the model in real-world deployment environments.

[0090] This step begins by selecting

Validate Model

Temporary Validation

Create Validation

[0091] During the validation process, attention should be paid to the confidence threshold configuration of the model output; it is generally recommended to set it to [value missing]. This is to filter out false positives with low confidence. (Coordinates of the annotation box) It must meet a minimum pixel requirement; for example, at a resolution of 2 megapixels, the bounding box should be no smaller than [a certain size]. The resolution of the verification image needs to be within a certain range of pixels to ensure the model can effectively recognize the target. to Between, and the recommended aspect ratio is To match the input standards of edge devices.

[0092] This step is widely used in scenarios such as compliance inspection in the kitchens of chain stores, personnel identification, and security monitoring. For example, in a kitchen setting, by uploading images of employees who were not trained and were not wearing masks or hats, the model's ability to identify abnormal behavior is verified, and the model parameters are adjusted based on the verification results. To improve recognition accuracy. Verification results. It can be used to evaluate the actual performance of the model after deployment on edge devices, providing data support for subsequent model iterations.

[0093] From a technical perspective, this step involves introducing scene images that were not used in the training process. This effectively tests the model's generalization ability on new data, preventing overfitting on the training data. Simultaneously, based on the validation results... The parameter adjustment mechanism dynamically optimizes the model's detection performance, improving its real-time performance and accuracy when deployed at the edge. This method acts as a bridge in the cloud-edge collaborative architecture, ensuring that the model's deployment effect at the edge is consistent with the training objectives in the cloud, thereby achieving efficient deployment and continuous optimization of AI capabilities.

[0094] The cloud-edge collaborative AI-powered intelligent analysis and management method for chain stores in this invention introduces a model verification step. It uses a temporary verification mode to upload scene images that were not used in training to evaluate model performance and dynamically adjusts model parameters based on detection confidence and bounding box coordinates. This further improves the adaptability and recognition accuracy of the AI ​​model in the changing scenarios of chain stores, thereby enhancing the overall intelligent analysis capability and management efficiency of the system.

[0095] Example 2 This invention addresses the operational management needs of chain stores by constructing a cloud-edge collaborative AI intelligent analysis and management system. It integrates core capabilities such as AI algorithm deployment, intelligent task configuration, anomaly identification, and closed-loop processing, adapting to multiple scenarios including kitchen compliance, management, and personnel identification. Through unified cloud control and local analysis on edge devices, it achieves intelligent monitoring and efficient management of store operations.

[0096] like Figure 2 As shown, the architecture adopts a three-tiered collaborative design of "cloud-edge-device": the AI ​​platform, as the algorithm hub, is responsible for the training, deployment, synchronization, and unified model library management of algorithm models, providing an intelligent capability foundation for the entire system; the cloud management platform undertakes global control functions, realizing remote deployment of algorithms to edge devices, unified task configuration distribution, and end-to-end data statistical analysis; the edge device layer consists of heterogeneous hardware such as IPCs, NVRs, and AI boxes, relying on local computing power to complete real-time video stream analysis, event detection, and data preprocessing, and reporting structured results to the cloud; the store scenario layer directly faces business applications, covering vertical scenarios such as kitchen compliance monitoring, facial recognition access control, and security anomaly early warning. Through layered decoupling and capability collaboration between cloud, edge, and device, the overall architecture ensures unified cloud management of edge nodes and efficient algorithm iteration, while fully leveraging the low-latency advantage of edge-side analysis, significantly reducing backbone network bandwidth pressure, adapting to the multi-scenario intelligent needs of chain stores, and achieving the dual goals of analysis efficiency and network optimization.

[0097] like Figure 3 As shown, the business configuration process follows the path of "model preparation - deployment decision - task definition": First, account configuration is completed on the AI ​​platform to establish a cloud authentication channel, and the target algorithm model is synchronized to the cloud model repository for centralized management; second, the deployment mode (cloud analysis or edge device local analysis) is selected according to the real-time requirements and resource constraints of the business, and the model is deployed to the edge nodes through a secure channel; finally, intelligent analysis tasks are created on the cloud management platform, and the detection area, event triggering rules and confidence thresholds are finely configured for the monitoring scenario to achieve flexible definition and policy control of tasks.

[0098] Furthermore, the AI ​​algorithm management module consists of an algorithm model library, an account configuration unit, and a model synchronization / deployment component. Its core function is to establish a connection with the AI ​​platform, enabling centralized cloud management and on-demand edge distribution of published algorithms (such as personnel wear detection and object recognition). This module supports dual cloud / edge deployment modes, allowing flexible selection of analysis locations based on business real-time requirements and network conditions. It also provides an algorithm tag filtering mechanism, subscribing to only relevant algorithms to reduce resource consumption and redundant computation on the edge side. The management process follows a standard four-step method: first, configure the AI ​​platform account (AppKey / Secret) to establish a secure authentication channel; second, synchronize the target algorithm model to the cloud model repository; third, select the deployment mode (cloud analysis or edge local analysis) according to task requirements; and finally, fine-tune the algorithm tags and recognition thresholds (adjustable from 30% to 100%) to ensure a balance between recognition accuracy and false alarm rate.

[0099] The intelligent analysis task module includes a task configuration engine, a visual detection area drawing tool, and rule setting components, designed to create highly customized intelligent analysis tasks, covering both timed image capture and real-time video analysis modes. Users can flexibly define detection areas for monitoring screens and set multi-dimensional trigger rules (such as target duration, quantity thresholds, behavioral logic, etc.) to accurately adapt to different scenario needs such as kitchen compliance and security monitoring. The key features of this module are reflected in two aspects: first, differentiated control of image capture frequency, with a minimum image capture interval of 5 minutes in the cloud and 1 minute at the edge, balancing cloud load and real-time performance; second, support for batch deployment capabilities, which can efficiently deploy the same task strategy to multiple edge devices simultaneously, enabling rapid replication and unified management of large-scale applications in chain stores.

[0100] The multi-scenario AI recognition module integrates scene-adaptive algorithms, face database management, and anomaly alarm components, comprehensively covering core business scenarios of chain stores and enabling flexible binding of algorithm capabilities with business needs. In kitchen management scenarios, it can accurately detect employee compliance (mask, hat, and uniform wearing recognition); in personnel management scenarios, it supports real-time customer flow statistics and movement analysis; in security monitoring scenarios, it provides proactive defense capabilities such as area intrusion detection and personnel exceeding limits alarm. Through pre-set scene templates and a dynamic algorithm loading mechanism, the module enables edge devices to quickly switch recognition tasks. Combined with the face database's blacklist / whitelist management and real-time comparison function, once an abnormal rule is triggered, it immediately notifies the responsible person through multi-level alarm channels (platform push, audible and visual alarms), forming a closed-loop response of "recognition-analysis-early warning".

[0101] In summary, the present invention can achieve the following technical effects: The cloud-edge collaborative architecture design adopts a four-level architecture of "AI platform - cloud management platform - edge device - store scenario", supports dual modes of centralized cloud management and local edge analysis, adapts to different bandwidth and scenario requirements, and takes into account both real-time performance and cost optimization.

[0102] The end-to-end AI algorithm management capability enables full-link control over algorithm synchronization, deployment, configuration, and optimization. It supports customizable algorithms to focus on, adjustment of recognition thresholds (30%-100%), and adaptation to multiple types of AI models (person wear detection, object recognition, etc.).

[0103] The scenario-based intelligent task configuration supports two task types: timed image capture and real-time video analysis. It allows you to define a dedicated detection area and set trigger rules (duration / quantity thresholds). The minimum analysis interval for a single channel is as low as 1 minute (edge) / 5 minutes (cloud).

[0104] The innovative design of the cloud-edge dual deployment dynamic adaptation mechanism algorithm allows for switchable deployment modes between the cloud and the edge. Edge devices perform real-time analysis and data reporting locally, while the cloud is responsible for centralized management and algorithm iteration, solving the pain points of high bandwidth consumption and high cloud analysis latency in high-traffic scenarios (protection points: deployment mode switching logic and data collaboration mechanism).

[0105] Customized detection rules and area configurations support precise delineation of detection areas (rectangle / polygon) and configuration of multi-dimensional trigger rules (time / quantity / threshold), adapting to the personalized monitoring needs of different store scenarios and solving the problem of poor adaptability of general AI analysis (protection point: flexible configuration technology of detection rules and areas).

[0106] The following description, in conjunction with the accompanying drawings, details the AI-powered intelligent analysis and management method for chain stores based on cloud-edge collaboration, according to an embodiment of the present invention: like Figure 4 As shown, the specific process of AI application is as follows: AI implementation follows a standardized process of five key steps: algorithm selection, data preparation, model training, model deployment, and model application. The algorithm selection phase involves matching the optimal algorithm template type (e.g., object detection, behavior recognition) to the business scenario. The data preparation phase involves collecting massive amounts of real-world images or video footage using cameras deployed in real-world scenarios to ensure the training data's suitability for the specific environment. The model training phase involves uploading the collected data to the AI ​​platform, using manual or automated intelligent annotation tools to accurately label the identified targets, selecting the application type, and initiating the training task to ultimately generate a high-precision algorithm model. The model deployment phase involves downloading the trained model files and distributing them in batches to edge devices equipped with GPU chips through the management platform, or directly importing them into the cloud inference platform. Simultaneously, IED intelligent analysis rule attributes (e.g., detection area, confidence threshold, trigger logic) are configured according to specific scenario requirements. The model application phase involves the deployed AI devices analyzing the video stream in real time and generating alarm events. Event information is pushed to industry application systems via the platform, allowing users to view complete alarm details, captured images, and associated recordings on the application, forming a complete closed loop from algorithm development to business value realization.

[0107] The algorithm selection is shown in Table 1: Table 1

[0108] Furthermore, data preparation, as the core foundation of the five major processes for AI implementation, directly determines the upper limit of deep learning algorithm performance in terms of its quality and scale, and is a key bottleneck in the quality of model performance. To ensure the recognition accuracy and robustness of the algorithm in real-world application scenarios, the collected materials must strictly adhere to the principle of homogeneity. This principle includes two dimensions: scene homogeneity requires that training materials must come from real deployment environments, accurately reproducing the lighting conditions, spatial layout, target distribution, and behavioral patterns of the target scene, and prohibiting the use of simulated or approximate scene data; imaging homogeneity emphasizes that the imaging parameters of the training data must be completely consistent with those of the actual application, including using the same model or specification of camera equipment, maintaining the same image resolution and aspect ratio, and reproducing the same installation height and shooting angle. In principle, real-time images of actual locations should be collected first, and the use of web-scraped images or photos taken casually by mobile phones as training materials is strictly prohibited. Violating the homology requirement will lead to overfitting or insufficient generalization ability of the model on both the training and test sets. This results in a "theoretical" phenomenon where the algorithm performs well in the laboratory environment but its accuracy drops sharply and the false positive rate soars when deployed in real-world scenarios. Ultimately, this prevents AI applications from meeting commercial delivery standards. Figure 5 As shown.

[0109] Furthermore, the richness of the training data is a core element determining the model's generalization ability, and it is necessary to ensure that the training data fully covers the diverse characteristics of real-world scenarios. First, the data should be readily available, prioritizing the use of existing surveillance footage or rapid accumulation through manual simulation to reduce data collection costs and timelines. Second, it must comprehensively cover the multi-dimensional changes of the target, including morphological differences (such as color, style, and structural variations), size variations (target size at different distances), angular diversity (top, side, and oblique views), occlusion (partial occlusion, overlapping occlusion), and spatiotemporal distribution (different time periods and different installation locations) to address the dynamic changes in the target's appearance in practical applications. Taking overflowing trash can recognition in residential settings as an example, given the diverse styles of trash cans commonly found in communities (fixed, mobile, flip-top; different colors such as red, blue, and green), separate data collection is required for each style. Simultaneously, considering the differences in viewing angles due to camera installation height and angle, coverage of different image locations and target scales is necessary. It is recommended to select multiple typical locations within the same community (including different styles and placements), collecting data from each location in three time periods: morning, noon, and afternoon. This ensures adequate coverage of lighting and pedestrian density variations. Each location should collect 300-500 high-definition images or 5-10 minutes of continuous video containing the target to construct a comprehensive and evenly distributed high-quality training set, preventing missed detections or false alarms due to limited sample size during actual deployment. Figure 6 As shown.

[0110] The pixel requirements for different model types are shown in Table 2: Table 2

[0111] Images of other resolutions need to be converted proportionally.

[0112] Furthermore, the data upload and annotation stage during model training serves as a bridge connecting the original materials and trainable data, and its level of refinement directly affects the model's convergence speed and recognition accuracy. The specific operational process is as follows: First, a dataset version is created on the AI ​​platform. Version management enables orderly traceability of data iterations and experimental comparisons. Each data expansion or cleaning generates an independent version number to ensure the reproducibility of the training process. Then, the optimal annotation method is selected based on the model type and target characteristics. Annotation box types include rectangular boxes (suitable for conventional target detection, such as standard geometric shapes like humans and vehicles), quadrilateral boxes (adapted to tilted or perspective-distorted targets, such as overturned trash cans or tilted trays), and polygonal boxes (used for high-precision semantic segmentation, such as irregular liquid spill areas or complex object outlines). The annotation process must strictly adhere to the consistency principle; the same annotation specification is used for the same target category across all samples to ensure the accuracy and completeness of the labels. This provides high-quality supervision signals for subsequent model training, avoiding model misjudgments or decreased generalization ability due to annotation noise.

[0113] The annotation box styles supported by different model types are shown in Table 3: Table 3

[0114] For object detection or hybrid models, rectangular annotations are usually sufficient. Quadrilateral annotations have higher requirements for device compatibility and computing power, and are suitable for some special scenarios, such as annotating densely packed targets or targets with large tilt angles. Quadrilateral annotations can better fit the target. Examples include pig counting, cargo box recognition, and rod-shaped objects, where dense targets or targets with large tilt angles are suitable for quadrilateral annotations. Figure 7 As shown.

[0115] Importing data allows you to choose between uploading images, folders (web or client), or compressed files (client): Web uploads are limited to 1000 images per session. For more than 1000 images, please compress them before uploading using the client. Image dimensions: 480*360 ≤ image resolution ≤ 4096*4096; single image size: ≤ 20MB; recommended aspect ratio: 16:9. Image formats: jpg, jpeg, png, bmp. Filename character restrictions: filenames cannot contain special characters: $, %, &.

[0116] After importing the data, you can begin annotation. Before annotation, you need to design an annotation plan. When annotating, it is recommended to pay attention to the following points: Clearly define the identification target: Accurately define the target to be identified (e.g., Figure 8 The belt misalignment function requires marking the idlers rather than the belt target, and determines the belt misalignment by identifying the number of idlers.

[0117] Labels should be flush with the target: The label box should be flush with the target, and should not be too large or too small (this type of problem often occurs and will affect the algorithm's performance), for example Figure 9 To ensure worker safety, it is necessary to inspect safety helmets (head and shoulder targets). The marking frame should be tightly fitted to the head and shoulder targets; for example... Figure 10 The labeled areas for targets with the same label must be consistent; for example... Figure 11 All targets meeting the above requirements must be labeled; omissions will affect algorithm performance. The label bounding box must be larger than the minimum pixel requirement: targets smaller than the minimum pixel requirement do not need to be labeled. Targets with more than 2 / 3 occlusion are not recommended for labeling. Figure 12 .

[0118] Model training is a core execution step in the AI ​​deployment process. After completing data annotation, you need to enter the "Train Model" module to create a training task. The specific operation path is as follows: Click the "Create Training" button to enter the training configuration interface, and then select the key parameters in sequence: First, select the pre-set basic model architecture (such as YOLOv8, ResNet, etc.), which should match the accuracy and speed requirements of the application scenario; second, specify the prepared training set and its specific version number to ensure data version traceability; then, filter the labels to be trained, supporting multi-label joint training or single-label specialized optimization; finally, select the application type according to the business objective (such as object detection, image classification, instance segmentation). In advanced configuration, you can enable high-precision training mode to extend the training cycle for a higher mAP metric, or enable intelligent tuning to automatically optimize hyperparameter combinations. After confirming that all configurations are correct, click "Start Training." The system will automatically allocate computing resources to start distributed training. During training, you can monitor the loss function convergence curve, mAP change trend, and training logs in real time. Upon completion, the system will automatically generate a model version and output an accuracy evaluation report.

[0119] Table 4 shows a comparison between supervised training and self-directed learning: Table 4

[0120] If the label is grayed out during training, it cannot be selected—regardless of the label level, it must be labeled.

[0121] After model training, a minimum data threshold must be met, meaning at least 40 effectively labeled images per class. This ensures the model has sufficient feature learning capabilities and avoids overfitting. To further verify the model's robustness and accuracy in real-world scenarios, the AI ​​platform includes a built-in model validation function, offering two methods for flexible evaluation: temporary validation and test set validation. The former supports rapid single or batch image validation, suitable for immediate feedback during model tuning. The latter provides systematic evaluation through a pre-set independent test set, outputting complete quantitative metrics such as confusion matrix, mAP, precision / recall. Taking the more commonly used temporary validation as an example, the process is as follows: Click on

Validate Model

Temporary Validation

Create Validation

Start Validation

[0122] It is recommended to select images that were not used in training in real-world application scenarios for verification; images without verification results are considered unidentified targets.

[0123] like Figure 13 As shown, based on the deployment method of the algorithm model and its corresponding deployment areas in the cloud, edge, and terminal, this solution can be divided into three complementary deployment modes to meet the differentiated needs of different scenarios for real-time performance, computing resources, and management granularity: Edge device deployment mode deploys algorithm models directly on AI platform edge devices (such as AI supercomputers, intelligent IPCs, and edge gateways) close to the data source through industry application platforms, relying on the local computing power of the devices for real-time inference and analysis. This mode pushes computing tasks down to the network edge, minimizing data transmission latency and achieving millisecond-level real-time response, making it particularly suitable for scenarios with extremely high immediacy requirements (such as security intrusion detection and kitchen compliance monitoring). Edge computing architecture not only ensures business continuity (it can still operate independently even if the network is interrupted), but also effectively reduces backhaul bandwidth pressure and cloud computing load, making it the solution that best embodies the "nearby computing" advantage in cloud-edge collaborative architectures.

[0124] The centralized deployment solution unifies core algorithms and powerful computing power at a regional or group center, forming a centralized intelligent analysis hub. Through an algorithm scheduling platform, various intelligent algorithms in the algorithm repository are uniformly managed and dynamically orchestrated, and then distributed to the central computing resource pool (mainly composed of high-performance GPU server clusters) on demand for large-scale parallel recognition and analysis. Based on scenario-based algorithm fusion technology, this solution unifies the scheduling of algorithm models for different business scenarios (such as customer flow statistics, security monitoring, and kitchen management), and achieves intelligent allocation and load balancing of GPU resources at the computing power layer, achieving deep intelligent integration of algorithms and computing power. The centralized deployment model facilitates unified control and version iteration of algorithm strategies across the entire enterprise headquarters, is suitable for medium- to large-scale chain deployments, balances real-time performance and management efficiency, and ensures that core data is processed locally in a closed loop.

[0125] The cloud deployment solution deploys the algorithm model on a public or private cloud platform, building Algorithm as a Service (AlaaS) capabilities. Front-end devices (such as cameras and sensors) upload keyframe images or structured data to the cloud platform. The cloud platform calls the corresponding model through standardized algorithm interfaces to perform cloud-based inference analysis and visualizes the recognition results in the management portal. The biggest advantage of this model lies in its openness and elastic scalability—it not only supports the enterprise's own business calls but also opens up algorithm capabilities to ecosystem partners or third-party applications through API interfaces, enabling cross-organizational and cross-system algorithm service sharing. In addition, the cloud platform can dynamically scale up and down computing resources according to business load, reducing initial investment costs. It is suitable for scenarios with lightweight access, unified management across multiple regions, or those requiring rapid trial and error innovation, building the "brain" role in the cloud-edge collaborative architecture.

[0126] These three deployment methods can be flexibly combined according to the enterprise's current IT infrastructure status, data security requirements, real-time needs and investment budget to form a hybrid architecture of "real-time edge response + unified central management and control + elastic cloud services", maximizing the technical value of cloud-edge collaboration.

[0127] like Figure 14As shown, this solution adopts a "cloud-edge-device" collaborative topology architecture, using an industry application platform as a unified management and control hub to achieve full lifecycle management of algorithm models and intelligent analysis tasks. Specifically, the Hikvision AI Open Platform, as the algorithm source, synchronizes the trained and optimized algorithm model to the industry application platform. The application platform manages model versions and configures task rules based on business needs, and distributes model files and analysis strategies in batches to Hikvision AI Open Platform devices (such as AI Super Brain, intelligent IPCs, and edge gateways) on various edge sides through a secure channel. The devices process video streams locally in real time, performing intelligent analysis tasks such as target detection and behavior recognition. The generated structured alarm information (including event screenshots, confidence levels, and timestamps) is instantly transmitted back to the application platform for visualization and centralized processing. At the same time, the system automatically collects positive samples and false alarm materials and feeds them back to the AI ​​training platform, forming a "data-driven, continuous iteration" algorithm optimization closed loop. This edge deployment solution is highly flexible, supporting differentiated configuration of analysis strategies based on stores, regions, or device types, achieving refined operations with a "one store, one policy" approach. The solution boasts significant advantages: leveraging edge computing to achieve edge perception and real-time response, it transforms raw video streams into structured event data for uploading, significantly reducing backbone network bandwidth consumption and cloud storage costs; the application platform provides one-click algorithm deployment capabilities, supporting batch distribution, version rollback, and canary upgrades, improving algorithm iteration efficiency by over 70%; with its modular design and scenario-based templates, the solution demonstrates excellent scenario adaptability and configuration flexibility, quickly adapting to diverse business needs such as kitchen compliance, customer flow statistics, and security monitoring, meeting the scaled and differentiated intelligent upgrade demands of chain businesses.

[0128] like Figure 15As shown, this solution adopts a centralized cloud deployment model, synchronizing the algorithm model trained on the AI ​​platform to the cloud platform and deploying it on a cloud server cluster to build a cloud-edge collaborative intelligent analysis architecture. The workflow is as follows: front-end devices (cameras, sensors) collect on-site images in real time and upload the captured images to the cloud platform as needed; the cloud platform calls the algorithm inference interface to perform cloud-based recognition and analysis, and displays the detection results in a visual manner on the management portal; at the same time, it supports local pre-analysis at the edge before uploading alarm images and structured data to the cloud platform for unified display; in addition, the cloud platform opens standard API interfaces, allowing third-party systems to call algorithm services and obtain inference results through the interfaces, realizing open capabilities; alarm images stored on the cloud platform can be automatically sent back to the AI ​​platform as optimization material, forming a closed loop of continuous iteration. The core components of the solution include: a cloud platform as the brain, which is responsible for the full lifecycle management of AI models (configuration, distribution, version control), unified reception and storage of AI alarm events, closed-loop processing of alarm events and work order flow, and model iteration and optimization driven by data feedback; cloud-based analytics is suitable for scenarios with large-scale system deployment and low real-time requirements, and obtains results by elastically calling inference interfaces, leveraging the scale advantage of cloud computing power; edge intelligent device analytics is designed for scenarios with high real-time requirements and high flexibility, and completes low-latency responses at the edge. The solution offers significant advantages: In terms of performance, it supports the deployment of high-precision, complex models, offering superior inference performance and accuracy. Leveraging a cloud-native architecture, it enables high-concurrency calls and elastic scaling of computing resources, easily handling business peaks. In terms of cost, cloud-based analytics can utilize existing monitoring equipment, avoiding redundant investment. It also supports pure cloud or cloud-edge hybrid deployments, eliminating the need for additional central servers and significantly reducing initial construction costs. In terms of management, the cloud platform provides unified and centralized management capabilities, allowing for rapid network access for store devices, precisely matching the speed of business expansion. It also supports dual-terminal viewing on PCs and mobile devices, enabling remote operation and maintenance anytime, anywhere. In terms of optimization, it supports an automatic data feedback mechanism, driving continuous optimization and iteration of AI model performance. In terms of business closed-loop, it deeply integrates with the inspection engine, automatically associating AI recognition results with inspection tasks and linking various inspection tools to achieve a complete business closed loop of "discovery-early warning-handling-verification," significantly improving inspection efficiency and management level.

[0129] like Figure 16As shown, this solution uses "convergence, integration, and empowerment" as its core value chain to build a central hub for intelligent video analysis, achieving unified convergence and governance, centralized intelligent analysis, dynamic intelligent scheduling, and open sharing of capabilities for multi-source heterogeneous video resources. In terms of technical implementation, the solution completes a closed loop through three mechanisms: First, unified algorithm management relies on an algorithm repository to uniformly manage perception, cognition, and decision-making algorithms, providing full lifecycle management capabilities from algorithm registration, version control, containerization, canary release to retirement, ensuring that algorithm assets are controllable and traceable; second, unified computing power scheduling integrates the central GPU cluster, edge computing nodes, and cloud elastic resource pools, achieving intelligent matching and dynamic allocation of algorithms and computing power based on multi-dimensional strategies such as task priority, real-time requirements, and resource load; finally, unified task management centrally configures, batches, monitors the status of, and retries abnormal intelligent analysis tasks across the entire network, supporting on-demand start / stop and resource preemption. The advantage of this architecture lies in achieving deep collaboration between video aggregation and governance and unified algorithm management. It can access video streams from devices with different manufacturers and protocols for standardized processing. At the same time, through scene algorithm fusion and intelligent orchestration and scheduling, it supports the flexible combination of multiple algorithm pipelines (such as "human detection + attribute recognition + behavior analysis" cascading) and automatically expands and shrinks according to business peaks and valleys, forming a flexible, efficient, and open intelligent video analysis capability platform that empowers diverse scenarios such as security, operations, and inspection.

[0130] like Figure 17 As shown, the IED (Intelligent Event Detect) module is a key logic processing layer connecting algorithm detection capabilities with business application requirements. Its core function is to transform the raw recognition results (such as target coordinates, category, and confidence level) output by the underlying AI algorithm into high-value events with business semantics. This module, through a visual rule configuration engine, allows users to define multi-dimensional composite logic rules flexibly without code development: in terms of counting logic, it can set area population statistics, vehicle entry and exit counts, and item dwell time; in terms of alarm logic, it supports alarm rules such as boundary crossing detection, area intrusion, exceeding personnel gathering limits, and lack of compliant attire, and allows customization of trigger thresholds (such as target dwell time, quantity limits, and confidence level thresholds); in terms of temporal logic, it can program behavioral sequence recognition of "detecting A first, then B" (such as moving items without scanning a code); in terms of combinational logic, it supports the fusion judgment of multiple algorithm results (such as triggering a safety violation event by linking "human presence + no safety helmet"). The flexibility of IED rules is reflected in the dynamic adjustment of parameters (such as polygon drawing of detection area, configuration of effective time period, target filtering conditions), support for blacklist and whitelist mechanism and rule priority scheduling, which decouples algorithm capabilities from business scenarios, significantly reduces the cost of customized development, and enables intelligent analysis needs in complex scenarios such as chain stores and industrial production to respond and iterate quickly.

[0131] IED rule classification and selection are shown in Table 5: Table 5

[0132] Taking the detection of anomalies in regional targets as an example, the main parameters are introduced as shown in Table 6: Table 6

[0133] like Figure 18 As shown, the AI ​​application's algorithm orchestration adopts a visual workflow design paradigm, flexibly constructing complex intelligent analysis logic through a node-based "start-perception-processing-end" chain. The start node, as the sole data entry point for the algorithm application, marks the input of a single image signal and the initiation of the perception and analysis process. Its core configuration item is the video processing frame rate, which directly determines the algorithm's sampling density of the video stream. A higher frame rate results in better spatiotemporal continuity but also higher computational consumption and fewer concurrent paths, requiring a trade-off between real-time business requirements and hardware resources. If the input source is discrete image data, this configuration can be ignored. The end node handles the final result output configuration, allowing users to select the desired result from any perception or processing node in the orchestration diagram as the application output. However, the current version does not support direct output of perception results from personnel information recognition models (such as faces and human figures) and moving target recognition; format conversion or feature aggregation must be performed through intermediate processing nodes. Perception capability is the core asset of the one-stop training platform, encompassing user-trained customized models and platform-pre-built open capabilities, endowing the orchestration process with the ability to recognize and understand targets (people, vehicles, objects, and behaviors) in the scene. Basic perception capability specifically refers to the ability for users to drag and drop their self-trained and published models (such as kitchen compliance detection and overflowing trash can detection) as basic nodes onto the orchestration canvas, freely combining them with the platform's general models (such as human detection and vehicle capture) to form a multi-algorithm pipeline. This enables complex business logic such as "detecting humans first and then recognizing workpieces" or "parallel operation of multiple target detections," supporting the rapid construction and iteration of scenario-based AI applications. As shown in Table 7: Table 7

[0134] The configurable parameters for the basic sensing node are shown in Table 8: Table 8

[0135] In addition to basic sensing capabilities, Hikvision's publicly available general-purpose models can also be selected as sensing nodes for orchestration. The currently supported auxiliary sensing capabilities and their perceptible targets and attributes are shown in Table 9. Table 9

[0136] In addition to basic and auxiliary perception capabilities, the platform supports selecting open capability models as perception nodes for orchestration. See the Open Capability Introduction for detailed capability information.

[0137] Target relationship processing is a core advanced function of the IED rule engine, designed to accurately identify complex business logic through spatial topology analysis and solve advanced scenario requirements such as "target coexistence" and "behavioral association". This module allows users to configure association rules for cross-category targets based on the intersection-over-union (IoU), thereby extracting deep semantic information such as "human-object interaction" and "object-object occlusion" from the "islands" of single detection results.

[0138] (1) Processing target: Determine the main target category that the current node needs to process and output, such as "personnel not wearing safety helmets" or "vehicles occupying fire lanes". This target will be used as the benchmark object for spatial relationship calculation.

[0139] (2) Target Relationship: The spatial relationship logic between targets is constructed through visual condition configuration, expressed as the paradigm of "overlapping with [another type of target] [more / less] [threshold percentage]". The calculation method of the overlap ratio provides three strategies to adapt to different scenario forms: the ratio of the smaller target area is suitable for judging whether the small target is completely contained by the large target (such as a mobile phone being held by a hand); the ratio of the larger target area is suitable for judging whether the large target has invaded the core area of ​​the small target (such as personnel entering the ROI warning zone); the ratio of the union area (standard IoU) is suitable for general target interaction judgment (such as the opening and closing state of the trash can lid and the body). The threshold setting needs to be combined with the target size and business tolerance. For example, in the kitchen compliance inspection, "the overlap between the mask and the face exceeds 85%" can be judged as wearing it correctly.

[0140] (3) Advanced Configuration: A time-dimensional anti-jitter mechanism is introduced to filter out instantaneous false triggers. The relationship duration defines the minimum duration for triggering an alarm (adjustable from 0 to 1800 seconds) to prevent false alarms caused by instantaneous interference such as birds and light shadows; the relationship duration sensitivity sets the trigger ratio within the time window (0-100%) to achieve a flexible judgment based on the "percentage of time the condition is met". The linkage logic between the two is as follows: within the set duration, if the cumulative time for the condition to be met reaches the duration × (1 - sensitivity), an alarm is triggered. For example, if the duration is configured as 10 seconds and the sensitivity is 90%, it means that within the 10-second time window, as long as the condition is met for a cumulative 1 second (10 × (1-90%)), an alarm can be triggered. The higher the sensitivity, the easier it is to trigger an alarm, which is suitable for scenarios with strict requirements for response speed; conversely, if the sensitivity is set to 30%, it needs to be met for a cumulative 7 seconds before triggering an alarm, which is suitable for compliance monitoring scenarios that require stable confirmation. This design effectively balances the timeliness and accuracy of alarms, avoiding the dilemma of missed alarms and false alarms.

[0141] When a water stain is on a book, its overlap with the book should be 100%; when the water stain is not on the book, the overlap should be 0%. Therefore, to detect water stains on a book, one could configure... Figure 19 As shown.

[0142] Similarly, we can apply this ability to other scenarios, such as: extracting the "text" on the "cow" from all the "text", or extracting the "moving" "construction vehicle" from all the "construction vehicles", etc.

[0143] Logical branches are the core control hub of the algorithm orchestration process, empowering users to dynamically schedule and orchestrate the perception process. This enables algorithm applications to overcome the limitations of linear execution and achieve layered decoupling and flexible orchestration of complex business logic. This module supports two modes: parallel branching and conditional judgment, to handle high-concurrency processing and intelligent routing decision-making scenarios, respectively.

[0144] Parallel branching mechanisms allow the simultaneous execution of two or more independent branches at runtime, enabling concurrent collaboration of multi-algorithm pipelines or heterogeneous processing. For example, two branches, "human detection + tooling recognition" and "regional intrusion detection," can run concurrently, each completing its analysis independently before subsequent nodes fuse and jointly assess the results. Alternatively, models of different precision (such as lightweight and high-precision models) can be called in parallel to perform layered processing on the same data source, balancing efficiency and accuracy. This mode is suitable for scenarios requiring simultaneous collection of multi-dimensional information or with ample computing resources. However, it should be noted that the number of branches is limited by the hardware's concurrent access capabilities and GPU memory capacity, resulting in linearly increasing resource consumption. Therefore, a reasonable assessment of device load is necessary.

[0145] The conditional judgment dynamically selects the execution path (Y branch or N branch) based on preset conditions, achieving intelligent routing and logical flow division. It supports configuring up to 8 independent conditions, which can be combined into complex condition groups through logical "AND / OR" relationships. The syntax format of condition configuration adopts a structured expression: [Target Type] Number of Targets (Confidence ≥ X%) (> / ≥ / < / ≤ / =) N. This design supports flexible combinations across categories, multiple thresholds, and quantity statistics, accurately characterizing business rules. For example, "Number of Humans (Confidence ≥ 80%) > 5" means that when there are more than 5 high-confidence human bodies in the screen, the condition is triggered, the Y branch executes "Crowd Gathering Alarm", and the N branch continues silent monitoring; or "Number of Trash Cans (Confidence ≥ 70%) and Number of Overflowing Trash (Confidence ≥ 85%) ≥ 1" triggers a sanitation warning. This mechanism enables the orchestration process to have environmental awareness and intelligent decision-making capabilities.

[0146] The advanced configuration introduces a time-dimensional anti-jitter mechanism to improve robustness: the condition duration setting trigger threshold must maintain a stable minimum length (0-1800 seconds) to prevent false triggers caused by momentary interference such as birds and light shadows; the condition duration sensitivity defines the effective trigger ratio within the time window (0-100%), and the two work together to achieve flexible judgment. The mathematical logic is: within the set duration, if the cumulative time the condition is met reaches duration × (1 - sensitivity), it is considered a trigger. For example, configuring a duration of 10 seconds and a sensitivity of 90% means that within a 10-second time window, triggering is achieved as long as the condition is met for a cumulative 1 second (10 × (1-90%)). Higher sensitivity makes it easier to trigger an alarm, suitable for security scenarios with stringent response speed requirements; if the sensitivity is set to 30%, it needs to be met for a cumulative 7 seconds before triggering, suitable for compliance monitoring scenarios requiring stable confirmation. This design effectively balances alarm timeliness and accuracy, avoiding the dilemma of missed and false alarms, and is particularly suitable for stable detection needs in complex environments such as changing lighting and temporary target occlusion, making the algorithm orchestration both flexible and reliable.

[0147] When the red light is on, Figure 20 As shown, because the classification model categorizes all drawn regions of interest or perceived objects during runtime, the classification results can sometimes be multiple. Therefore, in conditional judgments, the perception results of the classification model also have the concept of a target number. Based on the above, "when the scene is nighttime," when classifying the entire scene by default, the result is only one. Based on the conditional branching capability exemplified above, we can also apply it to other scenarios, such as: executing the logic for detecting people only when the image is clear, or using different models for perception based on whether the sky in the image is daytime or nighttime.

[0148] Example 3 To achieve the above embodiments, such as Figure 21As shown, this embodiment also provides a cloud-edge collaborative AI intelligent analysis and management device 10 for chain stores, including: The deployment mode configuration module 100 is used to dynamically configure the cloud deployment mode or edge device deployment mode of the AI ​​model according to the real-time business needs and network status of the chain stores. The video stream analysis module 200 is used to perform local real-time video stream analysis on the edge device based on the selected deployment mode, extract feature information of the target object and generate structured analysis results; The data transmission module 300 is used to upload the structured analysis results to the cloud management platform via a lightweight communication protocol, while retaining the original video data locally on the edge device. The task scheduling adjustment module 400 is used to dynamically adjust the model synchronization strategy and task scheduling rules between the cloud and the edge based on the operating status and analysis results fed back by the edge devices.

[0149] Furthermore, the deployment mode configuration module is also used for: Based on network state evaluation parameters and Deployment mode decision-making, among which Indicates the bandwidth usage threshold. Indicates the delay tolerance threshold; Classified by task priority and ,Will The task will be forcibly assigned to edge device deployment. Tasks are distributed to cloud or edge devices according to load balancing strategies.

[0150] Furthermore, the video stream analysis module is also used for: A confidence filtering mechanism is used to assess the confidence level of the detected target. Filter and keep only The target object; Perform a size filtering operation on the target bounding box size. To impose constraints, require pixels and Pixel.

[0151] Furthermore, the data transmission module is also used for: Use MQTT or gRPC protocols for structured data transmission, with packet size not exceeding [a certain value]. ; Through incremental synchronization mechanism Upload only the differences in the model, combined with breakpoint resume technology. Ensure transmission integrity.

[0152] The cloud-edge collaborative AI intelligent analysis and management device for chain stores in this invention introduces a model verification step. It uses a temporary verification mode to upload scene images that were not used in training to evaluate model performance and dynamically adjusts model parameters based on detection confidence and bounding box coordinates. This further improves the adaptability and recognition accuracy of the AI ​​model in the changing scenarios of chain stores, thereby enhancing the overall intelligent analysis capability and management efficiency of the system.

[0153] In the description of this specification, the references to "one embodiment," "some embodiments," "example," "specific example," or "some examples" refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0154] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two or three, unless otherwise explicitly specified.

Claims

1. A cloud-edge collaboration-based chain store AI intelligent analysis management method, characterized in that, Comprise: S1, according to the real-time business demand and network state of chain stores, dynamically configure the cloud deployment mode or edge device deployment mode of AI model; S2, based on the selected deployment mode, perform local real-time video stream analysis on the edge device, extract the feature information of the target object and generate structured analysis results; S3, upload the structured analysis results to the cloud management platform through a lightweight communication protocol, while retaining the original video data locally on the edge device; S4, according to the running state and analysis results fed back by the edge device, dynamically adjust the model synchronization strategy and task scheduling rules of the cloud and the edge.

2. The method of claim 1, wherein, The S1 further comprises: S11, based on network status evaluation parameters and make a deployment mode decision, wherein denotes a bandwidth occupation threshold value, denotes a delay tolerance threshold value; S12, Classified by task priority and ,Will Tasks will be forcibly assigned to edge devices for deployment. Tasks are distributed to cloud or edge devices according to load balancing strategies.

3. The method of claim 1, wherein, The S2 further comprises: S21, using a confidence filtering mechanism, the confidence of the detected target performing screening, only retaining the target object; S22, performing a size filtering operation on the target frame size constraints are imposed, requiring pixels and pixels.

4. The method of claim 1, wherein, The S3 further comprises: S31, structured data transmission using MQTT or gRPC protocol, packet size does not exceed ; S32, by incremental synchronization mechanism Upload only model difference part, combined with breakpoint resume technology Guarantee transmission integrity.

5. The method of claim 1, wherein, The S4 further comprises: S41, running the model state according to the feedback of the edge device , dynamically adjusting the model synchronization strategy , when , the error rate in , triggering model version rollback S42, based on the task scheduling algorithm Reassigning the task priority of the cloud and edge, the algorithm input parameters include the current network state , device load , and real-time task requirements .

6. The method of claim 1, wherein, Further comprising: S5, performing a model verification step by temporarily verifying the model uploading scene pictures not involved in the training and based on the verification result adjusting the model parameters wherein the verification result contains a confidence of the detected object and the bounding box coordinates .

7. A cloud-edge collaboration-based chain store AI intelligent analysis management device, characterized in that, Comprise: A deployment mode configuration module for dynamically configuring the cloud deployment mode or edge device deployment mode of AI model according to the real-time business demand and network state of chain stores; A video stream analysis module for performing local real-time video stream analysis on the edge device based on the selected deployment mode, extracting the feature information of the target object and generating structured analysis results; A data transmission module for uploading the structured analysis results to the cloud management platform through a lightweight communication protocol, while retaining the original video data locally on the edge device; A task scheduling adjustment module for dynamically adjusting the model synchronization strategy and task scheduling rules of the cloud and the edge according to the running state and analysis results fed back by the edge device.

8. The apparatus of claim 7, wherein, The deployment mode configuration module is further used for: Based on network state evaluation parameters and make deployment mode decisions, wherein denotes a bandwidth occupancy threshold, denotes a delay tolerance threshold; Classified by task priority and ,Will Tasks will be forcibly assigned to edge devices for deployment. Tasks are distributed to cloud or edge devices according to load balancing strategies.

9. The apparatus of claim 7, wherein, The video stream analysis module is further used for: A confidence filtering mechanism is adopted to filter the confidence of the detected target The target object is filtered and only the target object with a confidence value greater than a preset threshold is kept ​ Performing a size filtering operation on the target box size Constraining, requiring Pixels and Pixels.

10. The apparatus of claim 7, wherein, The data transmission module is further used for: Structured data transfer using MQTT or gRPC protocol, packet size no more than ; By incremental synchronization mechanism Upload only the model difference part, combined with the breakpoint resume technology Guarantee transmission integrity.

Images