Network camera-oriented monitoring system and data processing method thereof

Abstract

The application relates to the technical field of artificial intelligence, and discloses a network camera-oriented monitoring system and a data processing method thereof, so that the system has self-learning capability and the overall performance is improved. The method comprises the following steps: an AI recognition server screens out C type images with A <= target confidence <= B and D type images with target confidence > B; a file application server counts the detection results of each AI recognition server, stores the C type images and the D type images in different data storage blocks respectively, and outputs the classification results of the D type images to a user for secondary confirmation; according to the secondary confirmation information of the user, E type images with other types of alarm types and F type images which are misjudged in non-other types are screened out; and then an AI training server iteratively updates a deep learning model according to the C, E and F type images.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a monitoring system for network cameras and its data processing method. Background Technology

[0002] Traditional monitoring systems rely on manual, real-time visual inspections, which have significant limitations: First, prolonged, high-intensity monitoring can easily lead to visual fatigue and decreased attention span, reducing the efficiency of anomaly detection; second, due to limitations in manpower and hardware resources, there is an upper limit to the number of cameras that can be connected to a single terminal, making it difficult to meet the full coverage requirements of large-scale scenarios. While deep learning-based intelligent monitoring systems achieve automated target detection, they face multiple challenges in practical applications: the algorithms lack adaptability to complex scenarios, requiring significant human resources for data collection, annotation, model training, and iterative optimization; furthermore, due to data bias or limited model generalization ability, the system has a high risk of false alarms and missed alarms, affecting the accuracy and reliability of monitoring. Summary of the Invention

[0003] The purpose of this invention is to disclose a monitoring system for network cameras and its data processing method, which integrates data acquisition, annotation, model training and updating into the system, enabling the system to have self-learning capabilities and improve overall performance.

[0004] To achieve the above objectives, the present invention proposes a monitoring system for network cameras, including a client and a server cluster, wherein the server cluster includes: an AI training server, a file application server and at least two AI recognition servers; Each of the AI ​​recognition servers is used to capture images from a group of network cameras using independent threads, and then to detect each captured image in sequence based on the current version of the deep learning model, and to filter out C-class images with A≤target confidence≤B and D-class images with target confidence>B; The file application server is used to collect the detection results of each AI recognition server and store C-class images and D-class images in different data storage blocks respectively; it outputs the classification results of D-class images to the user for secondary confirmation, and filters out E-class images whose alarm type is affected by other types and F-class images that are misjudged among the other types based on the user's secondary confirmation information; The AI ​​training server is used to iteratively update the deep learning model based on images of classes C, E, and F, and to notify each of the AI ​​recognition servers to perform update processing.

[0005] Preferably, after the training task starts, the AI ​​training server determines whether the sum of the number of images in the temporary dataset synthesized from C, E, and F class images and the current version dataset is greater than a set threshold. If not, all remaining images in the temporary dataset are used as the training set to update the deep learning model again. If yes, the merged temporary dataset is randomly divided into a training set and a validation set. The deep learning model is obtained based on the training set, and the mAP value of each image in the validation set is calculated. After deleting some images in the validation set according to a preset deletion rule based on the mAP value, it determines whether the sum of the number of remaining images in the temporary dataset is greater than the set threshold. If yes, the training set and validation set are re-divided to enter the next loop until the sum of the number of remaining images in the temporary dataset is less than or equal to the set threshold. Then, all remaining images in the temporary dataset are used as the training set to update the deep learning model again. Finally, the AI ​​recognition servers are notified to perform the update process. In the training and iteration process of the deep learning model, the integrated Grounding Dino multimodal large model is used to automatically label images of classes C, E and F.

[0006] Optionally, the deep learning model employs the YOLOv11 algorithm.

[0007] Optionally, in the YOLOv11 algorithm, the constraint condition of "the target is in motion" is added to the intrusion behavior detection, and the optical flow algorithm is used to realize the motion detection function in the field of view. Each pixel in the image is classified as a moving pixel or a stationary pixel. When the detected target is on a moving pixel, the detection result is retained; when the detected target is on a stationary pixel, the detection result is discarded.

[0008] Optionally, during the image capture process, any AI recognition server maintains a condition variable for each capture thread to ensure that the buffer retains at most one latest image from the same network camera. Simultaneously, for each network camera, the position where the image is inserted into the buffer is determined by the recorded capture time, calculated using the following formula: ; in, It is the polling cycle. Taking screenshots takes time. It is the buffer position value according to the first-in-first-out (FIFO) principle; This indicates the number of elements in the buffer at the current moment; and during the image sorting process in the buffer, the latest image is added one by one. This buffer is a maximum of [number of elements]. In a linear linked list of elements, when an element is inserted into the index position of the buffer, the elements at the original index position and the elements at the following index positions are all shifted one position to the right.

[0009] Optionally, the network camera is also used to set a binary mask image with the same resolution as the network camera when only a part of the field of view needs to be monitored. The white area represents the area of ​​interest to be monitored, and the black area represents the non-monitored area. Before YOLOv11 inference, YOLOv11 training, and Grounding Dino automatic annotation, the image is cropped according to the maximum bounding box of the binary mask image, and the non-monitored areas in the image are blacked out. The algorithm only processes the cropped sub-image.

[0010] Optionally, the AI ​​training server is also used to re-label the dataset in a traversal manner based on the Grounding Dino multimodal large model when recognizing that a user has added or deleted new alarm types in the next version of the dataset.

[0011] Optionally, the file application server is also used to obtain the missed reporting information submitted by the client. The missed reporting information includes the missed reporting time period and the device information of the missed network camera, so that the AI ​​training server can automatically capture the missed images from the network camera and add them to the temporary dataset according to the missed reporting information.

[0012] To achieve the above objectives, the present invention also discloses a data processing method for a network camera monitoring system, wherein the method is applied to the above system and includes at least: Each AI recognition server uses an independent thread to capture images from a group of network cameras, and then performs sequential detection on the captured images based on the current version of the deep learning model to filter out C-class images with A≤target confidence≤B and D-class images with target confidence>B. The file application server collects the detection results of each AI recognition server and stores C-class and D-class images in different data storage blocks. It also outputs the classification results of the D-class images to the user for secondary confirmation and filters out E-class images whose alarm type is affected by other types and F-class images that are misjudged among the other types based on the user's secondary confirmation information. The AI ​​training server iteratively updates the deep learning model based on images of classes C, E, and F, and notifies each of the AI ​​recognition servers to perform the update process.

[0013] The present invention has the following beneficial effects: 1. By integrating data collection, annotation, model training, and updating into the system, the servers in the cluster have clear divisions of labor and simple and reliable collaboration logic, enabling the system to have self-learning capabilities and improving the overall performance of the system.

[0014] 2. The detection results of the current version of the deep learning model are classified according to confidence level. Among the alarm events pushed to the client operators, images of category E, marked as other types of alarms, are likely false alarms from the current model. Adding them to the new dataset will improve the adaptability of the new model to new / complex / changing scenarios, thereby reducing the false alarm rate. Meanwhile, some of the images of category C detected by the current model are correctly detected. Adding these to the new dataset will increase the diversity of positive samples in the new dataset, which will help improve the detection performance of the new model.

[0015] 3. When formulating deletion rules, generally, the higher the mAP value, the closer the model's detection results are to the true results. Adding such images to subsequent new datasets will not increase the sample diversity of the new dataset, nor will it improve the detection performance of the new model; instead, it will increase the training time of the new model. Therefore, such images can be deleted. Similarly, images whose classification results are consistent with those confirmed by humans are not included in the temporary dataset for the same purpose. For example, the specific deletion rule is to delete images in the top 10% of mAP and images with mAP > 0.98. If the ratio of the dataset and the validation set is 50% each, it can be guaranteed that at least 5% of the images in the temporary dataset will be deleted in one iteration, thus ensuring that the number of images in the temporary dataset will be less than or equal to the set threshold within a limited number of iterations.

[0016] 4. After confirming that the temporary dataset meets the expected threshold conditions, it is used entirely for model training, ensuring that information from all images is effectively focused, further improving the reliability and accuracy of the model.

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a distribution diagram of the constituent nodes of the monitoring system for network cameras disclosed in Embodiment 1 of the present invention.

[0019] Figure 2 This is a schematic flowchart of a data processing method for a network camera monitoring system disclosed in Embodiment 2 of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0021] Example 1 This embodiment discloses a monitoring system for network cameras, which is essentially a self-learning intelligent monitoring system based on human feedback. It fully utilizes the system characteristics and the powerful zero-shot generalization capability of the Grounding Dino multimodal large model to integrate data collection, annotation, model training, and updating operations into the system. This enables the system to have self-learning capabilities, greatly reducing the workload of R&D and maintenance personnel. At the same time, it enables the system to quickly adapt to complex scenarios and has the potential for widespread deployment, reducing the risk of false alarms and missed alarms and improving the overall performance of the system.

[0022] I. System Topology.

[0023] The system presents a star topology, such as Figure 1 As shown, the server cluster can be located on the same local area network (LAN) and use the LAN for data transmission and communication.

[0024] The system consists of network cameras, an AI recognition server, an AI training server, a file application server, an NVR (Network Video Recorder), and client applications. The network cameras support HTTP-based image capture, allowing for automatic capture of real-time images via a program. The AI ​​recognition server contains multiple AI inference cards (e.g., Huawei Ascend 300I PRO) and runs AI recognition programs. The AI ​​training server contains one or more AI training cards (e.g., Huawei Ascend 300I DUO) and runs AI training programs. The NVR automatically stores video stream data from all network cameras in real time. The file application server runs database management and business processing programs. The client application runs a web-based visualization program, providing a human-computer interaction interface.

[0025] II. System Workflow.

[0026] Assume the system has a total of M network cameras and N AI recognition servers, where M is typically much larger than N, for example, M=1000 and N=5. Assume all AI recognition servers have identical configurations, then each AI recognition server is responsible for monitoring the field of view of approximately M / N network cameras. The AI ​​recognition server runs an AI recognition program to identify moving targets such as people and vehicles within the field of view (assuming a target recognition confidence threshold of A, which can be 0.05), and sends the image, network camera number, and target information to the file application server. The business processing program merges the results of multiple AI recognition programs and records them in the database. When an image contains a target with a confidence level greater than B (B>A, which can be 0.3), the image and related information are pushed to the client interface, forming an alarm event. The client operator needs to perform a secondary confirmation for each alarm event to verify the actual alarm type, classifying the alarm event as an intrusion alarm, personnel disturbance, vehicle disturbance, animal disturbance, or other type of disturbance. The secondary confirmation result is simultaneously recorded in the database. If an intrusion alarm is triggered, the system can dispatch personnel to the site for handling. In this invention, "other types" refers to types that are not given priority attention for the time being but are left for observation, so as to provide dataset samples when expanding to new types later. "Non-other types" refers to types that have been clearly given priority attention by the system, including the aforementioned intrusion alarms, personnel impact, vehicle impact, animal impact, etc.

[0027] III. AI Recognition Program Design.

[0028] The AI ​​recognition program runs on an AI recognition server, enabling the identification of moving targets such as people and vehicles. Specifically, it uses the YOLOv11 algorithm from deep learning technology for target detection. An optical flow algorithm is used for motion detection within the field of view, classifying each pixel in the image as either a moving pixel or a stationary pixel. When a target is detected on a moving pixel, the detection result is retained. When a target is detected on a stationary pixel, the detection result is discarded. The retained recognition results are the output of the AI ​​recognition program. Because intrusion behavior inevitably involves two movement processes—entering and leaving the monitored area—adding the constraint of "target in motion" to intrusion behavior recognition is reasonable. This will not affect the final alarm rate while reducing the false alarm rate for fixed objects such as fire hydrants, lampposts, and shadows.

[0029] Alarm latency is a crucial performance indicator for intelligent monitoring systems, typically requiring a latency of less than 2 seconds. For a given network camera, the alarm latency is the sum of the image capture time, target recognition waiting time, target recognition time, and network transmission time to the file application server and client. Image capture time is the most unstable and may fail, while target recognition waiting time is programmable. When an AI recognition server manages P network cameras and includes Q AI inference cards, the AI ​​recognition program running on that server needs to solve a multi-producer-multi-consumer problem consisting of P image capture threads and Q recognition threads to minimize the mean and variance of the alarm latency for these P network cameras, thereby increasing the probability of ensuring that the alarm latency for all network cameras is below 2 seconds. To address this, this embodiment discloses an improved semaphore algorithm that dynamically adjusts the target recognition waiting time based on the actual image capture time, resolving the multi-producer-multi-consumer problem.

[0030] Specifically, assume the AI ​​recognition server is configured with P network cameras (e.g., 200) and has Q AI inference cards (e.g., 4). The AI ​​recognition program uses P independent threads to capture images from each of the P network cameras and Q independent threads to execute the target recognition algorithm. The algorithm pseudocode is shown in Table 1 below. Two improvements are made to the original semaphore algorithm for the multi-producer-multi-consumer problem: First, a condition variable is maintained for each image capture thread to ensure that the buffer holds at most one latest image from a particular network camera, thereby reducing the target recognition waiting time and thus lowering the average alarm latency. Second, for a given network camera, the position where image data is inserted into the buffer is inferred from the recorded capture time, ensuring that the sum of the capture time and the waiting time is as stable as possible, thereby reducing the variance of the alarm latency.

[0031] The estimation method is shown in the following formula (1), where period is the polling period, which is determined by the target recognition time and the number of network cameras P. It is generally relatively stable, for example, taking a value of 2 seconds. captrue_time is the time spent capturing the image, index is the buffer position value, and len(buffer) represents the number of elements in the buffer at the current time.

[0032] Formula (1) The buffer is a linear linked list containing at most P elements. Inserting an element into the buffer at a given index means that all elements at that index and those after it must be shifted one position to the right. Popping the 0th element from the buffer means that all elements at the 1st position and those after it must be shifted one position to the left. Buffer elements may contain fields such as image data and network camera IDs.

[0033] IV. AI Training Program Design.

[0034] The AI ​​training program enables the generation of a new dataset, training and updating of the YOLOv11 model, with the aim of improving the detection performance of the YOLOv11 object detection algorithm. The program can be started regularly every week or month, or manually, without real-time requirements, and its execution does not affect the normal operation of business systems.

[0035] The AI ​​training program operates as follows: First, the program downloads images from the file application server that are manually labeled as other types of impacts in alarm events, and downloads images containing targets with confidence levels between [A, B], obtaining a new image set. Then, the Grounding Dino algorithm from multimodal large model technology is used to automatically label the new image set, obtaining a new dataset. This new dataset is then merged with the old dataset to form a temporary dataset. In this embodiment, the temporary dataset also includes images selected based on user confirmation information as other types of impacts and images that were misjudged from other types. If the number of images in the temporary dataset is no greater than a set threshold (e.g., 100,000), then the temporary dataset is the new dataset. All images in the new dataset are then used as the training set to update the deep learning model again, and the YOLOv11 training program is run to train the new object detection model. Finally, the new model is updated to all AI recognition servers. If the number of images in the temporary dataset is greater than the set threshold, a dataset reduction operation is performed to ensure that the number of images in the subsequently obtained new dataset is always limited, thus ensuring that the training time for the YOLOv11 model is always limited.

[0036] Preferably, the dataset reduction technique in this embodiment first randomly divides the temporary dataset into 50% as the training set and the remaining 50% as the validation set. Then, the YOLOv11 training program is run to obtain a temporary detection model. Next, the temporary model is used to evaluate its detection performance on the validation set, and images with the top 10% mAP (Mean Average Precision) values ​​and images with mAP > 0.98 are deleted. Finally, the training set and the validation set after deleting some images are merged into a new temporary dataset. A higher mAP value indicates that the detection results of the temporary model are closer to the true results. Adding such images to a subsequent new dataset will not increase the sample diversity of the subsequent new dataset, nor will it improve the detection performance of the subsequent new model; instead, it will increase the training time of the subsequent new model. Therefore, such images can be deleted. Deleting images with the top 10% mAP values ​​and images with mAP > 0.98 ensures that at least 5% of the images in the temporary dataset are deleted in each iteration, thus ensuring that the number of images in the temporary dataset will be less than or equal to a set threshold within a finite number of iterations. Once the loop iterates until the sum of the remaining images in the temporary dataset is less than or equal to the set threshold, all the remaining images in the temporary dataset are used as the training set to update the deep learning model again; finally, each AI recognition server is notified to perform the update process.

[0037] Images marked as other alarm types in the alarm events pushed to client operators are likely false alarms from the current model. Adding them to the new dataset will improve the model's adaptability to new, complex, and changing scenarios, thereby reducing the false alarm rate. A portion of the images in the current model's detection set containing targets with confidence levels between [A, B] are correctly detected. Adding these to the new dataset will increase the diversity of positive samples, thus improving the detection performance of the new model.

[0038] Grounding Dino is a high-performance, open-vocabulary, multimodal object detection model based on the transformer architecture. While its recognition speed is relatively slow and unsuitable for real-time object detection tasks, it boasts very high accuracy and precision for detecting common types such as people, vehicles, and animals. Therefore, in this embodiment, it is used as an automatic annotation tool, with its detection results serving as the standard answer for the new model, thereby achieving the knowledge distillation process.

[0039] In addition, the first version of the YOLOv11 model and the first version of the dataset were provided manually during the system deployment phase.

[0040] V. Other supplementary situations that can be optimized.

[0041] A. Defense zone map function.

[0042] When only a portion of the network camera's field of view needs to be monitored, a zone map function can be introduced. The zone map is a binary mask image with the same resolution as the network camera; white areas represent the area of ​​interest (the monitoring area), and black areas represent the non-monitored area. Before YOLOv11 recognition, YOLOv11 training, and Grounding Dino automatic annotation, the image is cropped according to the maximum bounding box of the zone map, and the non-monitored areas are blacked out. The algorithm only processes the cropped image, making fuller use of computational resources and further improving the system's overall alarm performance.

[0043] B. One-click add / delete function for alarm target types When it's necessary to add or remove target types for alarms, leveraging the powerful generalization capabilities of the Grounding Dino multimodal large model in open-set target detection, the system only needs to execute an AI training program once after adding or deleting target types. This automatically re-labels the accumulated dataset and updates the newly trained model to the AI ​​recognition server. For example, if a user initially only focuses on human intrusion behavior, then the target type is only "person." Later, if the user wants to focus on animal intrusion behavior, they can add the "animal" type with a single click on the client interface, thus simultaneously monitoring both human and animal intrusion behaviors.

[0044] C. Add a function to detect missed samples.

[0045] Intelligent surveillance systems typically undergo a gradual improvement process. During actual use or testing, if missed detections are found, information such as the missed time period and network camera number can be manually submitted on the client side. The system automatically downloads images from the corresponding time period from the NVR, imports them into the AI ​​training server, and executes the AI ​​training program, thereby improving the detection performance of the new model and reducing the risk of subsequent missed detections.

[0046] Example 2 Corresponding to Embodiment 1 above, this embodiment discloses a data processing method for a network camera monitoring system. The method is applied to the aforementioned system, such as... Figure 2 As shown, it includes at least: Step S1: Each AI recognition server uses an independent thread to capture images from a group of network cameras. Then, based on the current version of the deep learning model, it performs sequential detection on each captured image and filters out Class C images with A ≤ target confidence ≤ B and Class D images with target confidence > B.

[0047] Step S2: The file application server collects the detection results of each AI recognition server and stores the C-class and D-class images in different data storage blocks. The classification results of the D-class images are output to the user for secondary confirmation. Based on the user's secondary confirmation information, the E-class images with alarm types affected by other types and the F-class images that were misjudged among the other types are filtered out.

[0048] Step S3: The AI ​​training server iteratively updates the deep learning model based on the C, E, and F class images and notifies each AI recognition server to perform the update process.

[0049] Preferably, this step can be implemented as follows: After the training task starts, the AI ​​training server determines whether the sum of the number of images in the temporary dataset synthesized from the C, E, and F class images and the current version dataset is greater than a set threshold. If not, all remaining images in the temporary dataset are used as the training set to update the deep learning model again. If yes, the merged temporary dataset is randomly divided into a training set and a validation set. The deep learning model is updated based on the dataset, and the mAP value of each image in the validation set is calculated. After deleting some images in the validation set according to the preset deletion rules based on the mAP value, it is determined whether the sum of the number of remaining images in the temporary dataset is greater than the set threshold. If yes, the training set and validation set are re-divided to enter the next loop until the sum of the number of remaining images in the temporary dataset is less than or equal to the set threshold. Then, all remaining images in the temporary dataset are used as the training set to update the deep learning model again. Finally, each AI recognition server is notified to perform the update process.

[0050] Furthermore, during the training and iteration of the deep learning model, images of classes C, E, and F are automatically labeled based on the integrated Grounding Dino multimodal large model.

[0051] The other logic in this embodiment is the same as in embodiment 1. For example, during the image capture process, any AI recognition server maintains a condition variable for each image capture thread to ensure that the buffer retains at most one latest image from the same network camera. At the same time, for each network camera, the position of the image inserted into the buffer is inferred by the recorded image capture time. The calculation formula is as follows: ; in, It is the polling cycle. Taking screenshots takes time. It is the buffer position value according to the first-in-first-out (FIFO) principle; This indicates the number of elements in the buffer at the current moment; and during the image sorting process in the buffer, the latest image is added one by one. This buffer is a maximum of [number of elements]. In a linear linked list of elements, when an element is inserted into the index position of the buffer, the elements at the original index position and the elements at the following index positions are all shifted one position to the right.

[0052] Furthermore, this embodiment also includes: when the network camera only needs to monitor a part of the field of view, a binary mask image with the same resolution as the network camera is set, where the white area represents the area of ​​interest to be monitored and the black area represents the non-monitored area. Before YOLOv11 inference, YOLOv11 training and Grounding Dino automatic annotation, the image is cropped according to the maximum bounding box of the binary mask image, and the non-monitored area in the image is blacked out. The algorithm only processes the cropped sub-image.

[0053] Alternatively, the method in this embodiment may further include: the file application server obtaining the missed reporting information submitted by the client, the missed reporting information including the missed reporting period and the device information of the missed network camera, and in conjunction with this, the AI ​​training server automatically capturing the missed images from the network camera and adding them to the temporary dataset based on the missed reporting information.

[0054] In summary, the monitoring system and data processing method for network cameras disclosed in this embodiment have at least the following beneficial effects: 1. By integrating data collection, annotation, model training, and updating into the system, the servers in the cluster have clear divisions of labor and simple and reliable collaboration logic, enabling the system to have self-learning capabilities and improving the overall performance of the system.

[0055] 2. The current version of the deep learning model classifies alarm events based on confidence levels. Images classified as "other types" (E-class) in the alarm events pushed to customer service representatives are highly likely to be false alarms from the current model. Adding these to the new dataset will improve the new model's adaptability to new, complex, and changing scenarios, thereby reducing the false alarm rate. Conversely, some of the C-class images detected by the current model are correctly detected. Adding these to the new dataset will increase the diversity of positive samples in the new dataset, thus improving the detection performance of the new model.

[0056] 3. When formulating deletion rules, generally, the higher the mAP value, the closer the model's detection results are to the true results. Adding such images to subsequent new datasets will not increase the sample diversity of the new dataset, nor will it improve the detection performance of the new model; instead, it will increase the training time of the new model. Therefore, such images can be deleted. Similarly, images whose classification results are consistent with those confirmed by humans are not included in the temporary dataset for the same purpose. For example, the specific deletion rule is to delete images in the top 10% of mAP and images with mAP > 0.98. If the ratio of the dataset and the validation set is 50% each, it can be guaranteed that at least 5% of the images in the temporary dataset will be deleted in one iteration, thus ensuring that the number of images in the temporary dataset will be less than or equal to the set threshold within a limited number of iterations.

[0057] 4. After confirming that the temporary dataset meets the expected threshold conditions, it is used entirely for model training, ensuring that information from all images is effectively focused, further improving the reliability and accuracy of the model.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A monitoring system for network cameras, comprising a client and a server cluster, characterized in that, The server cluster includes: an AI training server, a file application server, and at least two AI recognition servers; Each of the AI ​​recognition servers is used to capture images from a group of network cameras using independent threads, and then to detect each captured image in sequence based on the current version of the deep learning model, and to filter out C-class images with A≤target confidence≤B and D-class images with target confidence>B; The file application server is used to collect the detection results of each AI recognition server and store C-class images and D-class images in different data storage blocks respectively; it outputs the classification results of D-class images to the user for secondary confirmation, and filters out E-class images whose alarm type is affected by other types and F-class images that are misjudged among the other types based on the user's secondary confirmation information; The AI ​​training server is used to iteratively update the deep learning model based on images of classes C, E, and F, and to notify each of the AI ​​recognition servers to perform update processing. During the image capture process, any AI recognition server maintains a condition variable for each capture thread to ensure that the buffer retains at most one latest image from the same network camera. Simultaneously, for each network camera, the position where the image is inserted into the buffer is inferred from the recorded capture time, calculated using the following formula: ； in, It is the polling cycle. Taking screenshots takes time. It is the buffer position value according to the first-in-first-out (FIFO) principle; This indicates the number of elements in the buffer at the current moment; and during the image sorting process in the buffer, the latest image is added one by one. This buffer is a maximum of [number of elements]. In a linear linked list of elements, when an element is inserted into the index position of the buffer, the elements at the original index position and the elements at the following index positions are all shifted one position to the right.

2. The monitoring system for network cameras according to claim 1, characterized in that, The AI ​​training server is used to determine, after the training task starts, whether the sum of the number of images in the temporary dataset synthesized from the C, E and F class images and the current version dataset is greater than a set threshold. If not, the remaining number of images in the temporary dataset is directly used as the training set to update the deep learning model again. If so, the merged temporary dataset is randomly divided into a training set and a validation set. A deep learning model is obtained based on the training set, and the mAP value of each image in the validation set is calculated. After deleting some images in the validation set according to the preset deletion rules based on the mAP value, it is then determined whether the sum of the number of remaining images in the temporary dataset is greater than the set threshold. If so, the training set and validation set are re-divided to enter the next loop until the sum of the number of remaining images in the temporary dataset is less than or equal to the set threshold. Then, all the remaining images in the temporary dataset are used as the training set to update the deep learning model again. In the training and iteration process of the deep learning model, the integrated Grounding Dino multimodal large model is used to automatically label images of classes C, E and F.

3. The monitoring system for network cameras according to claim 2, characterized in that, The deep learning model uses the YOLOv11 algorithm.

4. The monitoring system for network cameras according to claim 3, characterized in that, In the YOLOv11 algorithm, the constraint of "the target is in motion" is added to the intrusion behavior detection and the optical flow algorithm is used to realize the motion detection function within the field of view. Each pixel in the image is classified as a moving pixel or a stationary pixel. When the detected target is on a moving pixel, the detection result is retained; when the detected target is on a stationary pixel, the detection result is discarded.

5. The network camera monitoring system according to any one of claims 1-4, characterized in that, The network camera is also used to set a binary mask image with the same resolution as the network camera when only a part of the field of view needs to be monitored. The white area represents the area of ​​interest to be monitored, and the black area represents the non-monitored area. Before YOLOv11 inference, YOLOv11 training and Grounding Dino automatic annotation, the image is cropped according to the maximum bounding box of the binary mask image, and the non-monitored areas in the image are blacked out. The algorithm only processes the cropped sub-image.

6. The network camera monitoring system according to claim 5, characterized in that, The AI ​​training server is also used to re-label the dataset in a traversal manner based on the Grounding Dino multimodal large model when recognizing that a user has added or deleted new alarm types in the next version of the dataset.

7. The network camera monitoring system according to claim 5, characterized in that, The file application server is also used to obtain the missed reporting information submitted by the client. The missed reporting information includes the missed reporting period and the device information of the missed network camera, so that the AI ​​training server can automatically capture the missed images from the network camera and add them to the temporary dataset based on the missed reporting information.

8. A data processing method for a network camera surveillance system, characterized in that, The method, when applied to the system as described in any one of claims 1 to 7, includes at least: Each AI recognition server uses an independent thread to capture images from a group of network cameras, and then performs sequential detection on the captured images based on the current version of the deep learning model to filter out C-class images with A≤target confidence≤B and D-class images with target confidence>B. The file application server collects the detection results of each AI recognition server and stores C-class and D-class images in different data storage blocks. It also outputs the classification results of the D-class images to the user for secondary confirmation and filters out E-class images whose alarm type is affected by other types and F-class images that are misjudged among the other types based on the user's secondary confirmation information. The AI ​​training server iteratively updates the deep learning model based on images of classes C, E, and F, and notifies each of the AI ​​recognition servers to perform the update process.

Images