Data processing method and system for video monitoring AI intelligent identification equipment

Abstract

The invention discloses a data processing method and system for video monitoring AI intelligent identification equipment, and the method specifically comprises the steps: obtaining the real-time video data of the video monitoring AI intelligent identification equipment of all factory monitoring areas, and forming an environment feature vector according to the real-time video data; dividing all factory monitoring areas into a plurality of environment categories based on the environment feature vectors, and determining a central monitoring area of each environment category; based on the central monitoring area, training a universal recognition model for each environment category, and adjusting the universal recognition model by adopting a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environment category; and deploying the specific identification model to the video monitoring AI intelligent identification equipment of each corresponding factory monitoring area. According to the invention, the intelligent level and the identification precision of the factory video monitoring system are obviously improved, the resource distribution is optimized, and the operation cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a data processing method and system for video surveillance AI intelligent recognition equipment. Background Art

[0002] With the continuous advancement of artificial intelligence technology, the field of video surveillance is ushering in a profound change. Traditional video surveillance methods are highly dependent on manual real-time viewing and analysis. This model is not only inefficient, but also easily leads to the omission of key information due to human factors (such as fatigue and negligence), which seriously weakens the monitoring effect. To meet this challenge, the industry is actively seeking the integration of artificial intelligence technology, aiming to build an automated and intelligent video surveillance analysis system to greatly improve monitoring efficiency.

[0003] Especially in the factory environment, the importance of video surveillance is particularly prominent. There are usually multiple monitoring areas inside the factory, and each area faces its own unique monitoring needs and challenges due to its different functions. For example, the production line area emphasizes real-time monitoring of product quality and production efficiency, the warehouse area needs to accurately track the dynamics of material storage and in and out of the warehouse, the safety channel is related to the safety of personnel passage, and the office area focuses on the protection of property and information security. These diverse needs make it difficult for traditional video surveillance methods to fully meet them, highlighting the necessity and urgency of technology upgrades.

[0004] Although video surveillance AI intelligent recognition equipment has been put into use and has achieved a preliminary improvement in the level of monitoring automation, it still faces two core problems. First, the equipment generally uses a unified recognition model to process video data from different monitoring areas, ignoring the significant differences in environmental characteristics of each area, resulting in uneven recognition results and difficulty in achieving expectations. Second, in terms of resource allocation strategy, existing equipment lacks sufficient flexibility and cannot dynamically adjust the allocation of server resources based on the real-time needs of each monitoring area, which in turn causes problems such as waste of resources or inefficient task completion, affecting the overall monitoring effectiveness. Summary of the invention

[0005] The purpose of the present invention is to provide a data processing method and system for video surveillance AI intelligent recognition equipment, which significantly improves the intelligence level and recognition accuracy of the factory video surveillance system, while optimizing resource allocation and reducing operating costs to solve at least one of the above-mentioned prior art problems.

[0006] In a first aspect, the present invention provides a data processing method for a video surveillance AI intelligent recognition device, the method specifically comprising: Acquire real-time video data of the video surveillance AI intelligent recognition equipment in all factory monitoring areas, extract the light intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data, and form an environmental feature vector; Based on the environmental feature vector, all factory monitoring areas are divided into multiple environmental categories, and a central monitoring area of ​​each environmental category is determined, wherein the environmental categories include production line area, storage area, safety passage, industrial equipment area and office area; Based on the central monitoring area, a general recognition model is trained for each environmental category, and the general recognition model is adjusted by using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category; Deploy the specific recognition model to the video surveillance AI intelligent recognition device of each corresponding factory monitoring area, and determine the recognition result of each factory monitoring area through the specific recognition model; The confidence score of each specific recognition model is determined according to the recognition result, and each specific recognition model whose confidence score is lower than a preset confidence threshold is optimized through a reinforcement learning algorithm.

[0007] In a second aspect, the present invention provides a data processing system for a video surveillance AI intelligent recognition device, the system specifically comprising: The first data processing module is used to obtain real-time video data of the video surveillance AI intelligent recognition equipment in all factory monitoring areas, extract the light intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data, and form an environmental feature vector; A second data processing module is used to divide all factory monitoring areas into multiple environmental categories based on the environmental feature vector, and determine the central monitoring area of ​​each environmental category, wherein the environmental categories include production line area, storage area, safety passage, industrial equipment area and office area; A third data processing module is used to train a general recognition model for each environmental category based on the central monitoring area, and adjust the general recognition model using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category; A fourth data processing module is used to deploy the specific recognition model to the video surveillance AI intelligent recognition device of each corresponding factory monitoring area, and determine the recognition result of each factory monitoring area through the specific recognition model; The fifth data processing module is used to determine the confidence score of each specific recognition model according to the recognition result, and optimize each specific recognition model whose confidence score is lower than a preset confidence threshold through a reinforcement learning algorithm.

[0008] In a third aspect, the present invention provides a computer device comprising: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, it implements the data processing method of the video surveillance AI intelligent recognition device as described in any one of the above methods.

[0009] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the data processing method of a video surveillance AI intelligent recognition device as described in any one of the above methods is implemented.

[0010] Compared with the prior art, the present invention has at least one of the following technical effects: 1. The present invention significantly improves the intelligence level and recognition accuracy of the factory video surveillance system, while optimizing resource allocation and reducing operating costs.

[0011] 2. By extracting environmental feature vectors, the present invention can fully understand the actual environmental conditions of each monitored area, providing an accurate data basis for subsequent environmental classification and model training.

[0012] 3. The present invention realizes the reasonable classification of monitoring areas through cluster analysis algorithm, and at the same time determines the central monitoring area of ​​each category, providing representative data samples for subsequent model training.

[0013] 4. The present invention obtains a universal recognition model applicable to various environmental categories through data labeling and neural network model training in the central monitoring area, thereby improving the generalization ability of the model.

[0014] 5. The present invention adopts a transfer learning method to enable the model to be fine-tuned according to the specific needs of different monitoring areas, thereby improving the accuracy and adaptability of the model.

[0015] 6. Through confidence scoring and reinforcement learning algorithms, the present invention can automatically identify and optimize models with poor performance, and continuously improve the overall recognition capability of the monitoring system.

[0016] 7. The present invention adopts genetic algorithm to realize dynamic optimization allocation of resources by comprehensively considering real-time demand and historical data, thereby improving resource utilization efficiency and ensuring efficient completion of tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 It is a flowchart of a data processing method of a video surveillance AI intelligent recognition device provided by an embodiment of the present invention; Figure 2 It is a structural diagram of a data processing system of a video surveillance AI intelligent recognition device provided by an embodiment of the present invention; Figure 3 It is a structural schematic diagram of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0020] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0021] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0022] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0023] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0024] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0025] In the embodiment of the present application, the execution subject of the process includes a terminal device, which includes but is not limited to: a server, a computer, a smart phone, a tablet computer, and other devices capable of executing the method disclosed in the present application. Figure 1 A flow chart of a data processing method of a video surveillance AI intelligent recognition device disclosed in an embodiment of the present invention is shown, and the details are as follows: S101, obtaining real-time video data of the video surveillance AI intelligent recognition equipment of all factory monitoring areas, extracting the light intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data, and forming an environmental feature vector.

[0026] In this embodiment, multiple video surveillance AI intelligent recognition devices are installed in the factory, which capture and transmit video data to the central processing system in real time. The central processing system receives and stores real-time video data from all monitoring areas through specific interfaces and protocols. The central processing system uses image processing technology to analyze the received real-time video data frame by frame. By extracting the pixel brightness information in the video frame and calculating the average value of these pixel brightness information, the system can obtain the average light intensity value of any factory monitoring area. This value can reflect the lighting conditions of the monitoring area and provide key data for the subsequent construction of environmental feature vectors. In order to obtain the shooting angle parameters, the system further analyzes the changing characteristics of the edge lines of objects in the video frame. By comparing the position and direction changes of the edge lines of objects in different frames, the system can infer the shooting angle of the camera. This step is crucial to understanding the perspective relationship and object shape of the monitoring screen. The system uses a Gaussian mixture model algorithm to perform background modeling on the video data. The algorithm can separate the foreground target and the static background in the video data. By calculating the color distribution degree and texture complexity of the static background, the system can obtain the background complexity of any factory monitoring area. Background complexity is an important indicator to measure the environmental complexity of the monitoring area, which is helpful for the subsequent construction of environmental feature vectors and model training. The extracted light intensity data, shooting angle parameters and background complexity are integrated to form environmental feature vectors. These feature vectors can fully reflect the environmental conditions of the factory monitoring area and provide basic data for subsequent environmental classification and model training.

[0027] S102, based on the environmental feature vector, all factory monitoring areas are divided into multiple environmental categories, and a central monitoring area of ​​each environmental category is determined, wherein the environmental categories include production line area, storage area, safety passage, industrial equipment area and office area.

[0028] In this embodiment, a suitable clustering algorithm (such as K-means clustering, DBSCAN clustering, etc.) is selected to perform cluster analysis on the environmental feature vector. According to the clustering results, the factory monitoring area is divided into multiple environmental categories, such as production line area, storage area, safety channel, industrial equipment area and office area. Each clustering result is assigned a corresponding environmental category label. For each environmental category, the mean or median of the environmental feature vectors of all its monitoring areas is calculated as the central feature vector of the category. According to the central feature vector, the monitoring area closest to the vector is located on the map or in the monitoring system as the central monitoring area of ​​the environmental category. The preliminarily determined central monitoring area is verified on the spot to ensure that it can represent the overall characteristics of the environmental category. According to the verification results, the central monitoring area is adjusted and optimized as necessary. By comparing the monitoring data under different environmental categories, the accuracy and practicality of the division results are evaluated, and user feedback is collected to understand whether the selection of the central monitoring area meets the actual needs. According to the evaluation results and user feedback, the division of environmental categories and the selection of the central monitoring area are optimized and adjusted. Clustering algorithms and center calculation methods are regularly updated and improved to adapt to changes in the factory environment and increased monitoring needs.

[0029] S103, based on the central monitoring area, a general recognition model is trained for each environmental category, and the general recognition model is adjusted by using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category.

[0030] In this embodiment, a large amount of annotated image or video data is collected from the central monitoring area. These data should cover various typical scenes and objects of the environment category. The collected data is preprocessed, including image enhancement, resizing, normalization, etc., to ensure data quality. A deep learning model suitable for the environment category is selected as the basic model, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). The basic model is trained using the preprocessed data to obtain a general recognition model. The model can recognize most objects and scenes in the environment category. The data of the central monitoring area is used as the source domain data, and the data of other factory monitoring areas is used as the target domain data. The transfer learning technology is used to transfer the knowledge in the source domain data to the target domain data. On the basis of the general recognition model, the data of each factory monitoring area is fine-tuned, including adjusting the model's weights, learning rates and other parameters to adapt to the data distribution and characteristics of different monitoring areas. Through iterative training and optimization, a specific recognition model is obtained for each factory monitoring area. These models can better adapt to the characteristics of their respective monitoring areas while maintaining general recognition capabilities. Use indicators such as accuracy, recall, and F1 score to evaluate general and specific recognition models, compare the performance of different models in different monitoring areas, and analyze the advantages and disadvantages of the models and the direction of improvement. Based on the evaluation results, optimize and adjust the model. This may include increasing the amount of data, improving data preprocessing methods, adjusting the model structure, etc. Update and maintain the model regularly to adapt to changes in the factory environment and increased monitoring needs.

[0031] In this embodiment, by training a general recognition model and fine-tuning it using transfer learning methods, specific recognition models suitable for different monitoring areas can be obtained, and these models can more accurately identify objects and scenes in their respective monitoring areas. Transfer learning technology allows us to fine-tune based on existing models instead of training from scratch, which greatly shortens the model training time and improves development efficiency.

[0032] S104, deploying the specific recognition model to the video surveillance AI intelligent recognition device of each corresponding factory monitoring area, and determining the recognition result of each factory monitoring area through the specific recognition model.

[0033] In this embodiment, the trained specific recognition model is deployed to the video surveillance AI intelligent recognition device of the corresponding factory monitoring area. The video surveillance AI intelligent recognition device captures the video data of the factory monitoring area in real time, and inputs the captured video data into the specific recognition model deployed in the device. The specific recognition model processes and analyzes the input video data to identify objects, scenes or behaviors in the video. The model outputs the recognition result according to the preset recognition rules and thresholds. The recognition result is output to the display interface of the device or the designated storage location in an appropriate form (such as text, image or sound, etc.). According to the recognition result, the corresponding alarm, linkage or other application scenarios are triggered.

[0034] In this embodiment, by deploying a specific recognition model, the video surveillance AI intelligent recognition device can more accurately identify objects, scenes or behaviors in the factory monitoring area, which helps to reduce false positives and false negatives and improve the accuracy and reliability of the monitoring system. By deploying a specific recognition model, the video surveillance AI intelligent recognition device can realize functions such as automatic recognition and alarm, which helps to improve the intelligence level of the factory and reduce the cost and risk of manual intervention.

[0035] S105, determining a confidence score of each specific recognition model according to the recognition result, and optimizing each specific recognition model whose confidence score is lower than a preset confidence threshold through a reinforcement learning algorithm.

[0036] In this embodiment, according to the application scenario and requirements of a specific recognition model, a suitable confidence evaluation index is selected. For example, in an object recognition task, the accuracy, recall rate or F1 score of the recognition result can be selected as a confidence evaluation index. The recognition results of each specific recognition model are evaluated, and the score is calculated according to the selected confidence evaluation index. The score is standardized to a confidence score between 0 and 1 for comparison and analysis. According to actual needs and application scenarios, a preset confidence threshold is set, which is used to determine whether the confidence of a specific recognition model meets the requirements. A reinforcement learning algorithm suitable for a specific recognition model optimization task is selected. For example, Q-learning, deep Q network (DQN), policy gradient method, etc. A specific recognition model with a confidence score lower than a preset confidence threshold is used as a model to be optimized. The optimization task is defined as improving the confidence score of the model to be optimized so that it reaches or exceeds the preset confidence threshold. The recognition results of the model to be optimized and the corresponding real label data are collected as training data for reinforcement learning. A reward function is designed to give rewards based on the confidence score of the model to be optimized and the accuracy of the recognition results. Use the reinforcement learning algorithm to train the model to be optimized, and improve the confidence score and recognition accuracy of the model through continuous trial and error and adjustment of model parameters. During the training process, regularly verify the performance of the model to be optimized to ensure that the model does not lose too much recognition accuracy while improving the confidence score. According to the verification results, make necessary adjustments and optimizations to the reinforcement learning algorithm and model parameters.

[0037] In this embodiment, the reinforcement learning algorithm is used to optimize the specific recognition model whose confidence score is lower than the preset confidence threshold, which effectively improves the confidence score of the model, helps to reduce false positives and false negatives, and improves the accuracy and reliability of the monitoring system. The reinforcement learning algorithm enables the model to better adapt to different monitoring scenarios and data distributions through continuous trial and error and adjustment of model parameters, enhances the generalization ability of the model, and improves the performance of the model in practical applications.

[0038] In some embodiments, in the above step S101, extracting the illumination intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data specifically includes: Extracting pixel brightness information from the real-time video data, and obtaining an average light intensity value of any factory monitoring area by calculating an average value of the pixel brightness information; By analyzing the change characteristics of the edge lines of objects in the real-time video data, the shooting angle parameters of any factory monitoring area are obtained; A Gaussian mixture model algorithm is used to perform background modeling on the real-time video data to separate the foreground target and the static background in the real-time video data. The background complexity of any factory monitoring area is obtained by calculating the color distribution and texture complexity of the static background.

[0039] In this embodiment, pixel brightness information of each frame image in the video data is extracted; for the pixel brightness information of each frame image, its average value is calculated to obtain the average light intensity value corresponding to the frame image; if the difference between the average light intensity value of the current frame and the average light intensity value of the previous frame is greater than a preset light intensity change threshold, it is determined that the lighting environment of the factory monitoring area has changed; according to the change trend of the average light intensity value of multiple consecutive frames of images, the Kalman filtering algorithm is used to filter the average light intensity value to remove noise interference in the light intensity data; by performing statistical analysis on the average light intensity value after filtering, the average light intensity level of the factory monitoring area over a period of time is obtained.

[0040] Perform edge detection on objects in the video to obtain edge line information of the objects; extract the change characteristics of the edge lines between different frames, and determine the shooting angle parameters of the current monitoring area based on the pre-established correspondence between the edge line change characteristics and the shooting angle.

[0041] The video data is transferred to the background modeling module for processing. In the background modeling module, the Gaussian mixture model algorithm is used to model the real-time video data, and the probability of each pixel belonging to the foreground or background is determined by the probability statistics method. According to the modeling results of the Gaussian mixture model algorithm, the pixels in the video data are divided into two parts: the foreground target and the static background, and the separation results of the foreground target and the static background are obtained. For the separated static background, its color distribution characteristics are calculated, and the frequency and proportion of different colors are counted to obtain the color distribution feature vector of the static background. For the separated static background, a texture feature extraction algorithm, such as grayscale co-occurrence matrix or local binary pattern, is used to extract the texture complexity characteristics of the static background. The color distribution feature vector and texture complexity characteristics of the static background are fused to construct a background complexity evaluation model, and the background complexity value of the factory monitoring area is obtained through model calculation.

[0042] For example, first, extracting pixel brightness information can help evaluate the lighting conditions in various areas of the factory. For example, in the furnace area of ​​a large steel plant, the average light intensity of the area can be obtained by calculating the average pixel brightness. The shooting angle parameters can be obtained by analyzing the changing characteristics of the edge line of the object, which has multiple meanings in factory monitoring. Taking the automobile assembly line as an example, by analyzing the edge line changes of the parts on the conveyor belt, the installation angle of the camera can be inferred. The Gaussian mixture model algorithm is widely used in background modeling. In a busy logistics warehouse, the algorithm can effectively separate the foreground targets such as moving workers and forklifts and the static backgrounds such as shelves and passages. By calculating the color distribution and texture complexity of the static background, the background complexity of the area can be obtained. In the reactor area of ​​a chemical plant, changes in light intensity may mean different stages of the reaction process. By continuously monitoring the average light intensity value, the operator can remotely judge the reaction state and adjust the parameters in time. At the same time, the acquisition of shooting angle parameters can help optimize the camera layout to ensure that the conditions inside the reactor can be clearly captured without being affected by equipment occlusion or light reflection. On the packaging line of a food processing plant, the calculation of background complexity can help identify abnormal situations. For example, if the background complexity of a certain area suddenly increases, it may mean that the packaging material is leaking or the product is piling up. The system can issue an alarm in time, which can not only improve production efficiency, but also reduce the risk of raw material waste and quality problems. Through the comprehensive application of these technologies, the factory monitoring system can operate more intelligently and accurately. Light intensity monitoring can ensure that the lighting conditions in the assembly area of ​​precision components are always in the best state; the acquisition of shooting angle parameters can optimize the visual system of the quality inspection link and improve the accuracy of defect detection; and background complexity analysis can help identify abnormal accumulation or personnel intrusion on the production line, thereby improving overall production efficiency and safety. By analyzing video data in real time, factory managers can obtain more comprehensive and in-depth insights into the production process, so as to make more informed decisions, optimize resource allocation, and improve production efficiency while ensuring worker safety and product quality.

[0043] In some embodiments, in the above step S102, based on the environmental feature vector, all factory monitoring areas are divided into multiple environmental categories, and the central monitoring area of ​​each environmental category is determined, specifically including: According to the distance measurement between the environmental feature vectors of each factory monitoring area, a cluster analysis algorithm is used to cluster all industrial monitoring areas, and industrial monitoring areas with similar distances and characteristics are divided into the same environmental category; For each environment category, calculate the mean vector of the environment feature vectors of all industrial monitoring areas within the environment category, and determine the mean vector as the central feature vector of the environment category; In each environment category, the industrial monitoring area that is closest to the central feature vector is determined as the central monitoring area.

[0044] In this embodiment, according to the distance metric between environmental feature vectors, the K-means clustering algorithm is used to cluster all industrial monitoring areas, and industrial monitoring areas with similar distances and similar features are divided into the same environmental category. For each environmental category, the mean vector of the environmental feature vectors of all industrial monitoring areas in the environmental category is calculated, and the mean vector is determined as the central feature vector of the environmental category. In each environmental category, the distance between the environmental feature vectors of each industrial monitoring area and the central feature vector of the category is calculated, and the industrial monitoring area with the closest distance is determined as the central monitoring area of ​​the environmental category. Real-time environmental monitoring data of each central monitoring area is obtained, environmental feature parameters are extracted, and the central feature vector of the corresponding environmental category is updated. The distance between the real-time environmental feature vector of each industrial monitoring area and the central feature vector of the environmental category to which it belongs is determined. If the distance exceeds a preset threshold, an early warning message is triggered, indicating that the environment of the monitoring area is abnormal. According to the historical environmental monitoring data of each industrial monitoring area, a support vector machine algorithm is used to establish a mapping model between environmental categories and environmental risk levels, and the environmental risk levels of each environmental category are evaluated in real time. Obtain high-risk environmental categories, adjust environmental parameter threshold settings for each industrial monitoring area, increase monitoring frequency, improve early warning sensitivity, and simultaneously control related environmental conditioning equipment to adjust and optimize environmental parameters to reduce environmental risks.

[0045] For example, a large integrated manufacturing plant is taken as an example. The plant contains multiple workshops of different natures, such as electronic component production, mechanical processing, and chemical processing. First, it is necessary to establish an environmental feature vector for each monitoring area. This vector may contain data in multiple dimensions such as light intensity, temperature, humidity, noise level, and air quality. In the electronic component production workshop, the environmental feature vector may show the characteristics of low noise, constant temperature and humidity, and high cleanliness. In contrast, the feature vector of the mechanical processing workshop may show characteristics such as high noise, large temperature fluctuations, and high metal dust content. The chemical processing workshop may show characteristics such as high concentration of special gases, high temperature, and strong corrosiveness. Next, cluster analysis algorithms such as K-means or hierarchical clustering are used to classify these feature vectors. The distance metric can use Euclidean distance or Manhattan distance. Through this method, the following clustering results may be obtained: precision manufacturing environment category (including electronic component production, precision instrument assembly, etc.), heavy industrial environment category (including mechanical processing, metal smelting, etc.), chemical processing environment category, etc. For each environment category, its central feature vector is calculated. For example, the central feature vector of the precision manufacturing environment category may show an average temperature of 22°C, a relative humidity of 45%, a noise level of 45 dB, a dust concentration of less than 0.1 μg / m3, etc. This central feature vector represents the typical environmental characteristics of the category and can be used as a reference standard for the category. After determining the central feature vector, the actual monitoring area closest to it can be found and set as the central monitoring area. This area can be used as a representative of the environmental category to quickly assess the overall status of the entire category. For example, in the precision manufacturing environment category, the integrated circuit packaging workshop with the highest cleanliness and the most stable temperature and humidity control may be selected as the central monitoring area. This method of clustering and central area determination has multiple meanings. First, it can help managers quickly understand the overall status of different types of environments in the factory without having to check all areas one by one. Second, it provides a benchmark for anomaly detection: if the environmental characteristics of an area are too different from the central feature vector of the category to which it belongs, it may mean that an abnormal situation has occurred and further investigation is required. Finally, this method can also guide resource allocation: equipment maintenance and upgrades in the central monitoring area can be prioritized because the status of these areas has the greatest impact on the monitoring effect of the entire category. Through this systematic environmental analysis and monitoring method, factories can manage complex production environments more efficiently and improve production efficiency, while ensuring that environmental parameters in each area are always maintained at optimal levels, thereby ensuring product quality and employee safety.

[0046] In some embodiments, in the above step S103, training a general recognition model for each environment category based on the central monitoring area specifically includes: Determine a monitoring recognition task for each central monitoring area, and annotate the first video data of the central monitoring area according to the monitoring recognition task to form a monitoring recognition data set, wherein the monitoring recognition task includes product defect recognition, worker operation behavior recognition, cargo classification location recognition, industrial equipment abnormality recognition, and outsider recognition; The monitoring identification data set is input into a neural network model for modeling training to obtain a general identification model of the environment category corresponding to each central monitoring area.

[0047] In this embodiment, the first video data of each central monitoring area is obtained, and the first video data is annotated according to the preset monitoring recognition task to obtain a monitoring recognition data set, wherein the monitoring recognition task includes product defect recognition, worker operation behavior recognition, cargo classification position recognition, industrial equipment abnormality recognition and outsider recognition. The monitoring recognition data set is input into the neural network model, and the neural network model is trained using a supervised learning algorithm to obtain a preliminarily trained neural network model. The environmental category information of the central monitoring area is obtained, and the preliminarily trained neural network model is fine-tuned according to the environmental category information to obtain a general recognition model corresponding to the environmental category of each central monitoring area. For each central monitoring area, real-time video data is collected as test data, and the test data is input into the corresponding general recognition model for prediction to determine whether there are abnormal situations such as product defects, improper worker operation behavior, wrong cargo placement, abnormal operation of industrial equipment and the appearance of outsiders. If the prediction result of the general recognition model shows that there is an abnormal situation, an early warning signal is triggered, and the relevant personnel are prompted to deal with the abnormal situation in time through the human-computer interaction interface. The processing result information of the abnormal situation is obtained, and the processing result information and the corresponding test data are added to the monitoring recognition data set to form new training data. Use new training data to retrain the general recognition model, continuously improve the model's generalization ability and recognition accuracy, and achieve adaptive optimization of monitoring and recognition tasks.

[0048] For example, taking the precision manufacturing environment category as an example, the central monitoring area may be an integrated circuit packaging workshop. The main monitoring and identification tasks of this workshop include product defect identification, worker operation behavior identification, and outsider identification. The product defect identification task involves detecting various defects that may occur in the packaging process, such as poor welding, packaging cracks, etc. In order to establish an effective recognition model, a large number of packaged products need to be video-shot and annotated. During the annotation process, professionals will carefully review the video frames of each product and mark the location, type, and severity of the defects. This fine annotation work provides a high-quality data foundation for subsequent model training. The worker operation behavior identification task aims to ensure that workers comply with standard operating procedures and improve production efficiency and safety. During the annotation process, it is necessary to classify and mark the various operating actions of workers, such as picking, placing, and debugging. This includes not only correct operating behaviors, but also some common wrong operations, so that the model can identify and warn in time. The outsider identification task is crucial to maintaining the safety and confidentiality of the workshop. When annotating, it is necessary to distinguish between normal staff and outsiders, including their clothing characteristics, behavior patterns, etc., which helps the system quickly identify people who enter sensitive areas without authorization. By inputting these labeled data sets into the neural network model for training, a general recognition model suitable for precision manufacturing environments can be obtained. This model can handle multiple recognition tasks at the same time, improving the efficiency and flexibility of the monitoring system. For example, the use of convolutional neural networks (CNN) combined with target detection algorithms can effectively identify product defects; while recurrent neural networks (RNN) are more suitable for analyzing the sequence of workers' operation behaviors. The model trained by this method has strong generalization ability and can be applied not only to the central monitoring area, but also to other areas of the same environmental category. This greatly improves the practicality and economic benefits of the model. For example, the model trained in the electronic component production workshop can be better applied to the precision instrument assembly workshop because they belong to the same environmental category and have similar environmental characteristics and monitoring requirements. This general recognition model method based on environmental categories not only improves the intelligence level of the monitoring system, but also provides a powerful tool for factory management. It can timely detect product quality problems, standardize worker operations, and ensure workshop safety, thereby comprehensively improving production efficiency and product quality. At the same time, this method also has good scalability. With the continuous accumulation of data and continuous optimization of the model, its recognition accuracy and scope of application will continue to improve, providing strong technical support for the intelligent transformation of the factory.

[0049] In some embodiments, in the above step S103, the general recognition model is adjusted by using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category, specifically including: Determine the attention weight distribution parameters of each factory monitoring area on different monitoring recognition tasks; Based on the attention weight distribution parameters, a transfer learning method is used to adjust the general recognition model of the environmental category corresponding to each factory monitoring area to obtain a specific recognition model for each factory monitoring area.

[0050] In this embodiment, the monitoring video data of each factory monitoring area is obtained, the video data is preprocessed, and the key frames and regions of interest are extracted. For different monitoring and recognition tasks of each factory monitoring area, such as personnel recognition, vehicle recognition, etc., the corresponding attention weight distribution parameters are determined. According to the attention weight distribution parameters, the preprocessed video data is feature extracted to obtain feature vectors representing different monitoring and recognition tasks. A transfer learning method is adopted, based on a pre-trained general recognition model, and the feature vector is input into the model for fine-tuning training to obtain a recognition model adapted to the specific environment of each factory monitoring area. The real-time video stream of each factory monitoring area is processed using the adjusted specific recognition model to realize real-time recognition and tracking of different monitoring objects. According to the recognition results, it is judged whether an abnormal situation occurs. If an abnormal situation occurs, an alarm mechanism is triggered to notify relevant personnel to handle it. The specific recognition model of each factory monitoring area is continuously optimized, and the newly collected data is added to the training set by incremental learning to continuously improve the generalization ability and recognition accuracy of the model.

[0051] For example, taking the precision manufacturing environment as an example, an integrated circuit packaging workshop may face multiple tasks such as product defect recognition, worker operation behavior recognition, and outsider recognition. Different weights can be assigned to these tasks according to the specific situation of the workshop. Assuming that the product quality problem in this workshop is more prominent, the weight of product defect recognition can be set to 0.5, worker operation behavior recognition to 0.3, and outsider recognition to 0.2. This weight distribution reflects the current main focus of the workshop and helps the model allocate computing resources more efficiently. After determining the attention weight distribution parameters, the general recognition model can be adjusted using the transfer learning method to adapt to the needs of a specific monitoring area. The core idea of ​​transfer learning is to transfer the knowledge learned on one task to another related task, thereby reducing the need for a large amount of labeled data for the new task. In this case, you can start with a pre-trained general recognition model that has been trained on a large amount of precision manufacturing environment data. Then, use a small amount of labeled data from a specific workshop to fine-tune the model. Specifically, a fine-tuning strategy can be adopted to keep most of the parameters of the pre-trained model unchanged and only adjust the last few layers. This method can retain the model's understanding of general features while adapting to the unique characteristics of a specific workshop. For example, for the task of product defect recognition, the product samples specific to the workshop can be used for fine-tuning so that the model can identify the defect types specific to the workshop. During the fine-tuning process, the attention mechanism can also be introduced to adjust the model's attention to different tasks according to the previously determined weight distribution parameters. For example, for the product defect recognition task with a higher weight, the model's attention at the relevant feature extraction layer can be enhanced to make it more sensitive to capture subtle changes in the product surface. In this way, a model that retains the general recognition ability and can adapt to the needs of a specific monitoring area can be obtained. This specific recognition model can more accurately identify the product defects of the workshop and better understand the workers' operating behavior, while not neglecting the safety-related task of identifying outsiders. This method based on attention weights and transfer learning has significant technical advantages. First, it greatly reduces the need for a large amount of labeled data, reducing the cost and time of model training. Second, by adjusting the attention weights, the model can quickly adapt to the focus tasks of the workshop at different times, improving the flexibility of the system. Finally, this method ensures that the model can respond accurately to specific environments while maintaining versatility, improving overall monitoring efficiency.

[0052] In some embodiments, in the above step S105, determining the confidence score of each specific recognition model according to the recognition result, and optimizing each specific recognition model whose confidence score is lower than a preset confidence threshold by a reinforcement learning algorithm specifically includes: Based on the recognition results, determining the accuracy of each specific recognition model on different monitoring recognition tasks; Determining a confidence score for each specific recognition model based on the accuracy rates on different monitoring recognition tasks and the attention weight distribution parameters; Comparing the confidence score of each specific recognition model with a preset confidence threshold, and determining a number of first specific recognition models having confidence scores lower than the preset confidence threshold score; Using the model parameter adjustment method of the first specific recognition model as the action space, using the model parameter information of the first specific recognition model as the state space, and using the confidence score of the first specific recognition model as the reward function; Based on the action space, the state space and the reward function, a Q-Learning algorithm is used to optimize each first specific recognition model.

[0053] In this embodiment, the recognition results of each specific recognition model on different monitoring recognition tasks are obtained; for the recognition results of each specific recognition model, its accuracy on each monitoring recognition task is counted; the pre-established attention weight distribution parameter is obtained; the accuracy of each specific recognition model on different monitoring recognition tasks is weighted and summed with the corresponding attention weight distribution parameter to obtain the comprehensive performance index of the model, that is, the confidence score. The confidence score is compared with the preset confidence threshold to determine whether the confidence score is lower than the preset confidence threshold. If the confidence score is lower than the preset confidence threshold, the specific recognition model is determined as the first specific recognition model. For each first specific recognition model, its model parameter adjustment method is obtained, and the model parameter adjustment method is used as the action space of reinforcement learning. For each first specific recognition model, its model parameter information is obtained, and the model parameter information is used as the state space of reinforcement learning. For each first specific recognition model, its confidence score is obtained, and the confidence score is used as the reward function of reinforcement learning. According to the obtained action space, state space and reward function, the Q-Learning algorithm is used to optimize each first specific recognition model to obtain an optimized first specific recognition model. For each optimized first specific recognition model, its confidence score is obtained, and the confidence score of the optimized first specific recognition model is compared with the preset confidence threshold to determine whether the confidence score of the optimized first specific recognition model is higher than the preset confidence threshold. If the confidence score of the optimized first specific recognition model is higher than the preset confidence threshold, the optimized first specific recognition model is determined as the target specific recognition model. All target specific recognition models are obtained, all target specific recognition models are combined into an optimized specific recognition model group, and the optimized specific recognition model group is output.

[0054] For example, taking the integrated circuit packaging workshop as an example, it is assumed that for the three tasks of product defect recognition, worker operation behavior recognition and outsider recognition, the accuracy rates of 85%, 78% and 92% were obtained respectively. These accuracy rates reflect the performance of the model on each task and provide a basis for subsequent optimization. Combining the accuracy rate and the attention weight distribution parameter, the confidence score of each specific recognition model can be calculated. For example, if the weight of product defect recognition is 0.5, the weight of worker operation behavior recognition is 0.3, and the weight of outsider recognition is 0.2, the overall confidence score of the model can be calculated by weighted average. Assume that the calculated confidence score is 83%. Comparing the calculated confidence score with the preset threshold, the model that needs to be optimized can be identified. If the preset threshold is 85%, the above model will be marked as the first specific recognition model that needs to be optimized. This method can effectively screen out models with poor performance and concentrate resources for optimization. For models that need to be optimized, the Q-Learning algorithm in reinforcement learning can be used for optimization. In this process, the model parameter adjustment method is defined as an action space, including adjusting the learning rate, changing the network structure, etc. The model parameter information constitutes the state space, which reflects the current state of the model. The confidence score is used as a reward function to guide the optimization direction. In specific implementation, a series of discrete parameter adjustment actions can be designed, such as adjusting the learning rate to 0.8 times or 1.2 times the current value. The state space can include key parameters such as the current number of network layers and the number of neurons in each layer. The reward function can be designed as the difference between the old and new confidence scores to encourage adjustments that improve model performance. Through multiple iterations, the Q-Learning algorithm can gradually find the optimal parameter adjustment strategy. For example, the algorithm may find that increasing the number of convolutional layers can significantly improve the accuracy of product defect recognition, thereby improving the overall confidence score. The advantage of this method is that it can automatically explore a large number of possible optimization directions and find the most effective optimization strategy. This optimization method based on reinforcement learning can not only improve model performance, but also adapt to the dynamically changing factory environment. As production conditions change, the model can continue to optimize and maintain efficient recognition capabilities. In addition, this method also provides valuable insights for factory managers to help understand the key factors that affect model performance, thereby improving production processes and quality control on a wider scale.

[0055] In some embodiments, in the above steps S101 to S105, the method further includes: Obtain real-time video traffic, model reasoning demand information, and historical resource usage data for each factory monitoring area; With minimizing resource utilization and maximizing task completion rate as optimization goals, a fitness function is set in combination with the real-time video traffic, the model reasoning requirement information and the historical resource usage data; A genetic algorithm is used to randomly generate a set of initial resource allocation schemes for each factory monitoring area. Based on the fitness function, selection operations, crossover operations and mutation operations are performed on the initial resource allocation schemes until the convergence condition of the fitness function or the preset number of iterations is reached, and a target resource allocation scheme is generated. The target resource allocation scheme includes an allocation ratio of computing resources and an allocation ratio of storage resources.

[0056] In this embodiment, the real-time video traffic data of each factory monitoring area is obtained, and the required computing resources and storage resources are determined according to the size of the video traffic and the model reasoning requirements. The historical resource usage data of each factory monitoring area is obtained, and the relationship between the resource utilization rate and the task completion rate is obtained by analyzing the historical data, and the fitness function is constructed to minimize the resource utilization rate and maximize the task completion rate as the optimization goal. A genetic algorithm is used to randomly generate a set of initial resource allocation schemes for each factory monitoring area, and each scheme includes the allocation ratio of computing resources and storage resources. According to the fitness function, the initial resource allocation scheme is evaluated and the fitness value of each scheme is calculated. The higher the fitness value, the better the scheme. The initial resource allocation scheme is selected, and the scheme with a high fitness value is more likely to be selected. Through the crossover operation, some genes of the selected high-quality scheme are exchanged to generate a new resource allocation scheme. The newly generated resource allocation scheme is mutated, that is, random perturbations are introduced into the genes of the scheme to maintain the diversity of the population and avoid the algorithm from falling into a local optimum. Determine whether the convergence condition of the fitness function or the preset number of iterations has been reached. If not, return to step 4 to continue iterative optimization. If reached, output the resource allocation scheme with the highest fitness value as the target scheme, including the computing resource and storage resource allocation ratio of each factory monitoring area.

[0057] Exemplarily, first, obtaining the real-time video traffic, model reasoning requirements and historical resource usage data of each monitoring area is the basis for optimization. For example, in a large automobile manufacturing plant, the assembly line, paint shop and quality inspection area may have different monitoring requirements. The assembly line may require high frame rate video to capture fast-moving parts, while the quality inspection area requires high-resolution images to detect tiny defects. Setting the fitness function is the key to the optimization process. The function should comprehensively consider resource utilization and task completion rate. In practical applications, resource utilization can be defined as the ratio of allocated resources to total available resources, and task completion rate can be the ratio of the number of successfully processed video frames to the total number of frames. The fitness function can be a weighted sum of these two indicators, and the weights can be adjusted according to the specific needs of the factory. The application of genetic algorithms makes the optimization of resource allocation schemes efficient and feasible. The generation of the initial scheme can be based on historical data or completely random. For example, 10% to 50% of the available computing resources and storage resources can be randomly allocated to each area. The selection operation can adopt the roulette method so that the scheme with high fitness has a greater probability of being selected. The crossover operation can exchange the resource allocation ratios in different schemes, such as exchanging the computing resource allocation ratio of the assembly line in scheme A with the corresponding value in scheme B. The mutation operation can randomly adjust the resource allocation of a certain area, such as increasing or decreasing the storage resources of the paint shop by 5%. The advantage of this optimization method is that it can adapt to the complex and changing factory environment. For example, when the factory introduces a new production line or updates equipment, the system can quickly adjust the resource allocation to adapt to the new demand. In addition, by analyzing the optimization results, factory managers can gain insight into the resource demand patterns of different areas and provide a basis for future equipment investment and system upgrades. In actual applications, it may be found that some areas have sudden high demands during specific periods. For example, the quality inspection area may need more computing resources to handle a large number of inspection tasks when the production batch is changed. By introducing the time dimension into the optimization model, the system can predict and adjust the resource allocation in advance to ensure good performance during peak periods. The target resource allocation scheme generated in the end includes not only static allocation ratios, but also dynamic adjustment strategies. For example, when a sudden increase in video traffic on the assembly line is detected, the system can automatically borrow resources from other temporarily idle areas. This flexible resource allocation mechanism can further improve the efficiency and reliability of the entire monitoring system. Through this genetic algorithm-based optimization method, factories can maximize the effectiveness of the monitoring system under limited resource conditions and improve production efficiency and product quality.

[0058] Reference Figure 2 An embodiment of the present invention provides a data processing system 2 for a video surveillance AI intelligent recognition device, and the system 2 specifically includes: The first data processing module 201 is used to obtain real-time video data of the video surveillance AI intelligent recognition equipment in all factory monitoring areas, extract the light intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data, and form an environmental feature vector; A second data processing module 202 is used to divide all factory monitoring areas into multiple environmental categories based on the environmental feature vectors, and determine the central monitoring area of ​​each environmental category, wherein the environmental categories include production line area, storage area, safety passage, industrial equipment area and office area; The third data processing module 203 is used to train a general recognition model for each environmental category based on the central monitoring area, and adjust the general recognition model using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category; The fourth data processing module 204 is used to deploy the specific recognition model to the video surveillance AI intelligent recognition device of each corresponding factory monitoring area, and determine the recognition result of each factory monitoring area through the specific recognition model; The fifth data processing module 205 is used to determine the confidence score of each specific recognition model according to the recognition result, and optimize each specific recognition model whose confidence score is lower than a preset confidence threshold through a reinforcement learning algorithm.

[0059] It is understandable that if Figure 1 The contents of the data processing method embodiment of the video surveillance AI intelligent recognition device shown in the figure are all applicable to the data processing system embodiment of the video surveillance AI intelligent recognition device. The functions specifically implemented by the data processing system embodiment of the video surveillance AI intelligent recognition device are similar to those of the embodiment shown in the figure. Figure 1 The data processing method embodiment of the video surveillance AI intelligent recognition device shown in the figure is the same as that of the embodiment of the video surveillance AI intelligent recognition device shown in the figure, and the beneficial effects achieved are the same as those of the embodiment of the video surveillance AI intelligent recognition device shown in the figure. Figure 1 The beneficial effects achieved by the data processing method embodiment of the video surveillance AI intelligent recognition device shown are also the same.

[0060] It should be noted that the information interaction, execution process and other contents between the above-mentioned systems are based on the same concept as the embodiment of the method of the present invention. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0061] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0062] Reference Figure 3 The embodiment of the present invention further provides a computer device 3, comprising: a memory 302 and a processor 301 and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, the data processing method of the video surveillance AI intelligent recognition device described in any one of the above methods is implemented.

[0063] The computer device 3 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will appreciate that Figure 3 It is only an example of computer device 3 and does not constitute a limitation on computer device 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components, for example, it may also include input and output devices, network access devices, etc.

[0064] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0065] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may also be an external storage device of the computer device 3, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device 3. Further, the memory 302 may also include both an internal storage unit of the computer device 3 and an external storage device. The memory 302 is used to store an operating system, an application program, a boot loader (Boot Loader), data, and other programs, such as the program code of the computer program, etc. The memory 302 may also be used to temporarily store data that has been output or is to be output.

[0066] An embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the data processing method of the video surveillance AI intelligent recognition device as described in any of the above methods is implemented.

[0067] In this embodiment, if the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, USB flash drive, mobile hard disk, disk or optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0068] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0069] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0070] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0071] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Claims

1. A data processing method for a video surveillance AI intelligent recognition device, characterized in that: The method specifically comprises: Acquire real-time video data of the video surveillance AI intelligent recognition equipment in all factory monitoring areas, extract the light intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data, and form an environmental feature vector; Based on the environmental feature vector, all factory monitoring areas are divided into multiple environmental categories, and a central monitoring area of ​​each environmental category is determined, wherein the environmental categories include production line area, storage area, safety passage, industrial equipment area and office area; Based on the central monitoring area, a general recognition model is trained for each environmental category, and the general recognition model is adjusted by using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category; Deploy the specific recognition model to the video surveillance AI intelligent recognition device of each corresponding factory monitoring area, and determine the recognition result of each factory monitoring area through the specific recognition model; The confidence score of each specific recognition model is determined according to the recognition result, and each specific recognition model whose confidence score is lower than a preset confidence threshold is optimized through a reinforcement learning algorithm.

2. The method according to claim 1, characterized in that The step of extracting the illumination intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data specifically includes: Extracting pixel brightness information from the real-time video data, and obtaining an average light intensity value of any factory monitoring area by calculating an average value of the pixel brightness information; By analyzing the change characteristics of the edge lines of objects in the real-time video data, the shooting angle parameters of any factory monitoring area are obtained; A Gaussian mixture model algorithm is used to perform background modeling on the real-time video data to separate the foreground target and the static background in the real-time video data. The background complexity of any factory monitoring area is obtained by calculating the color distribution and texture complexity of the static background.

3. The method according to claim 1, characterized in that Based on the environmental feature vector, all factory monitoring areas are divided into multiple environmental categories, and the central monitoring area of ​​each environmental category is determined, specifically including: According to the distance measurement between the environmental feature vectors of each factory monitoring area, a cluster analysis algorithm is used to cluster all industrial monitoring areas, and industrial monitoring areas with similar distances and characteristics are divided into the same environmental category; For each environment category, calculate the mean vector of the environment feature vectors of all industrial monitoring areas within the environment category, and determine the mean vector as the central feature vector of the environment category; In each environment category, the industrial monitoring area that is closest to the central feature vector is determined as the central monitoring area.

4. The method according to claim 1, characterized in that: The training of a general recognition model for each environment category based on the central monitoring area specifically includes: Determine a monitoring recognition task for each central monitoring area, and annotate the first video data of the central monitoring area according to the monitoring recognition task to form a monitoring recognition data set, wherein the monitoring recognition task includes product defect recognition, worker operation behavior recognition, cargo classification location recognition, industrial equipment abnormality recognition, and outsider recognition; The monitoring identification data set is input into a neural network model for modeling training to obtain a general identification model of the environment category corresponding to each central monitoring area.

5. The method according to claim 4, characterized in that The transfer learning method is used to adjust the general recognition model to obtain a specific recognition model for each factory monitoring area in each environmental category, specifically including: Determine the attention weight distribution parameters of each factory monitoring area on different monitoring recognition tasks; Based on the attention weight distribution parameters, a transfer learning method is used to adjust the general recognition model of the environmental category corresponding to each factory monitoring area to obtain a specific recognition model for each factory monitoring area.

6. The method according to claim 5, characterized in that Determining the confidence score of each specific recognition model according to the recognition result, and optimizing each specific recognition model whose confidence score is lower than a preset confidence threshold by a reinforcement learning algorithm, specifically includes: Based on the recognition results, determining the accuracy of each specific recognition model on different monitoring recognition tasks; Determining a confidence score for each specific recognition model based on the accuracy rates on different monitoring recognition tasks and the attention weight distribution parameters; Comparing the confidence score of each specific recognition model with a preset confidence threshold, and determining a number of first specific recognition models having confidence scores lower than the preset confidence threshold score; Using the model parameter adjustment method of the first specific recognition model as the action space, using the model parameter information of the first specific recognition model as the state space, and using the confidence score of the first specific recognition model as the reward function; Based on the action space, the state space and the reward function, a Q-Learning algorithm is used to optimize each first specific recognition model.

7. The method according to any one of claims 1 to 6, characterized in that The method further comprises: Obtain real-time video traffic, model reasoning demand information, and historical resource usage data for each factory monitoring area; With minimizing resource utilization and maximizing task completion rate as optimization goals, a fitness function is set in combination with the real-time video traffic, the model reasoning requirement information and the historical resource usage data; A genetic algorithm is used to randomly generate a set of initial resource allocation schemes for each factory monitoring area. Based on the fitness function, selection operations, crossover operations and mutation operations are performed on the initial resource allocation schemes until the convergence condition of the fitness function or the preset number of iterations is reached, and a target resource allocation scheme is generated. The target resource allocation scheme includes an allocation ratio of computing resources and an allocation ratio of storage resources.

8. A data processing system for video surveillance AI intelligent recognition equipment, characterized in that: The system specifically comprises: The first data processing module is used to obtain real-time video data of the video surveillance AI intelligent recognition equipment in all factory monitoring areas, extract the light intensity data, shooting angle parameters and background complexity of each factory monitoring area according to the real-time video data, and form an environmental feature vector; A second data processing module is used to divide all factory monitoring areas into multiple environmental categories based on the environmental feature vector, and determine the central monitoring area of ​​each environmental category, wherein the environmental categories include production line area, storage area, safety passage, industrial equipment area and office area; A third data processing module is used to train a general recognition model for each environmental category based on the central monitoring area, and adjust the general recognition model using a transfer learning method to obtain a specific recognition model for each factory monitoring area in each environmental category; A fourth data processing module is used to deploy the specific recognition model to the video surveillance AI intelligent recognition device of each corresponding factory monitoring area, and determine the recognition result of each factory monitoring area through the specific recognition model; The fifth data processing module is used to determine the confidence score of each specific recognition model according to the recognition result, and optimize each specific recognition model whose confidence score is lower than a preset confidence threshold through a reinforcement learning algorithm.

9. A computer device, characterized in that: include: A memory and a processor and a computer program stored in the memory, when the computer program is executed on the processor, implements the data processing method of the video surveillance AI intelligent recognition device as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by the processor, the data processing method of the video surveillance AI intelligent recognition device as described in any one of claims 1 to 7 is implemented.

Images