Intelligent camera monitoring system and method based on Internet of Things

Abstract

The invention relates to the technical field of the Internet of Things, and particularly discloses an intelligent camera monitoring system and method based on the Internet of Things. The system comprises a camera node, a region edge computing node and a cloud management platform. According to the invention, the local lightweight reasoning module is deployed at each camera node and the data desensitization and feature extraction are executed, the processing task of the original video data is sunk to the edge, and only the desensitized feature data or the lightweight decision result is uploaded, so that the data volume transmitted to the cloud is greatly reduced, the network bandwidth pressure is effectively relieved, and the user experience is improved. The operation and maintenance cost of the system is reduced, the way of exposing original privacy data on a transmission link is fundamentally cut off, and user privacy protection is remarkably enhanced.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, specifically relating to an IoT-based intelligent camera monitoring system and method. Background Technology

[0002] In the field of intelligent security and surveillance, the deep application of IoT technology is driving the development of traditional monitoring systems towards intelligence and networking. This system, by deploying a large number of camera nodes and combining them with artificial intelligence algorithms, aims to achieve core functions such as environmental perception, behavior recognition, and automatic early warning of abnormal events. Among these, IoT-based intelligent camera monitoring systems have become a key technological direction for improving public safety and regional management efficiency.

[0003] The IoT-based intelligent camera monitoring system aims to collect video data through distributed camera nodes and use artificial intelligence models for real-time analysis and decision-making. The core of this technology lies in achieving efficient and accurate environmental perception and anomaly detection, while ensuring the system's real-time performance and reliability.

[0004] Existing technologies typically employ a centralized architecture, uploading video data from all camera nodes to the cloud for unified storage and model training. This architecture faces significant challenges: the continuous transmission of massive amounts of video data puts enormous pressure on network bandwidth, increasing system deployment and maintenance costs; and the raw video data carries a high risk of user privacy breaches during transmission and centralized processing.

[0005] Uniformly trained models struggle to adapt to the unique environmental conditions of each camera node, such as changes in lighting and shooting angles. This leads to decreased generalization ability in real-world scenarios and long update cycles, hindering timely responses to dynamic changes in the local environment. Therefore, there is an urgent need for an intelligent monitoring solution that can effectively balance data processing efficiency, privacy protection, and model adaptability. Summary of the Invention

[0006] The technical problems this invention aims to solve are overcoming the network bandwidth pressure and high operation and maintenance costs caused by the continuous transmission of massive amounts of video data in existing centralized camera monitoring systems; the risk of user privacy leakage during the transmission and centralized processing of raw video data; and the problems of decreased model generalization ability and excessively long update cycles caused by the inability of a unified training model to adapt to the unique environmental conditions of each camera node. This invention provides an IoT-based intelligent camera monitoring system and method, aiming to construct a distributed intelligent processing architecture. Through localized data processing and personalized model adaptation mechanisms, it significantly reduces network load, enhances user privacy protection, and improves the accuracy and real-time response capability of the model in changing environments.

[0007] The technical solution of the present invention is to construct an intelligent camera monitoring system based on the Internet of Things. The system includes multiple camera nodes deployed in the monitoring area, regional edge computing nodes that communicate with the camera nodes, and a cloud management platform that communicates with the regional edge computing nodes. Multiple camera nodes form the perception layer, responsible for collecting raw video stream data. Regional edge computing nodes form the edge processing layer, responsible for receiving and processing data uploaded by their associated camera nodes. The cloud management platform forms the central coordination layer, responsible for system-level model management, policy distribution, and global status monitoring.

[0008] Furthermore, each camera node integrates a video acquisition module, a first preprocessing module, and a local lightweight inference module. The video acquisition module is configured to continuously capture the raw video stream of the monitored scene at a preset frame rate. The first preprocessing module is connected to the video acquisition module and is configured to perform data desensitization and feature extraction operations on the raw video stream. The data anonymization operation specifically includes real-time blurring of face and license plate regions in the video stream to generate anonymized video data. The feature extraction operation specifically includes using a pre-built primary feature extraction network to extract keyframes and their corresponding multi-dimensional feature vectors from the anonymized video data. These multi-dimensional feature vectors include scene semantic features, motion optical flow features, and temporal context features.

[0009] Furthermore, the local lightweight inference module is connected to the first preprocessing module, which is configured to load and run a lightweight anomaly detection model. This lightweight anomaly detection model receives the multi-dimensional feature vector output by the first preprocessing module as input and outputs a preliminary anomaly probability score. When the initial anomaly probability score exceeds the first preset threshold, the local lightweight inference module triggers an alarm and generates an edge decision data packet containing keyframes, multi-dimensional feature vectors, and an initial anomaly probability score. At the same time, the edge decision data packet is uploaded to the edge computing node of its region.

[0010] Furthermore, the regional edge computing node includes a data aggregation module, a model fine-tuning module, and a second decision module. The data aggregation module is configured to receive and cache edge decision data packets from all camera nodes under its jurisdiction. The model fine-tuning module is configured to perform a personalized model adaptation process for each camera node. The personalized model adaptation process is as follows: The model fine-tuning module obtains the basic anomaly detection model from the cloud management platform, and uses the historical feature vectors in the edge decision data packets recently generated by the camera node and the manual review labels to form a local training dataset. The basic anomaly detection model is incrementally trained to generate fine-tuning model parameters specific to the camera node, and the fine-tuning model parameters are sent to the local lightweight inference module of the corresponding camera node for model update.

[0011] Furthermore, the second decision-making module is configured to perform aggregation analysis and final decision-making. The second decision-making module receives edge decision data packets forwarded by the data aggregation module and runs a regional-level decision-making model with a complexity higher than the local lightweight inference model. The regional-level decision-making model comprehensively analyzes edge decision data packets from multiple camera nodes in the same area, performs spatiotemporal correlation analysis, and calculates the regional comprehensive risk index. When the regional comprehensive risk index exceeds a second preset threshold, the second decision-making module generates a high-level alarm signal and handling suggestions, and uploads the high-level alarm signal, handling suggestions, and related data summaries to the cloud management platform.

[0012] Furthermore, the cloud management platform includes a model repository module, a policy management module, and a global monitoring module. The model repository module stores multiple versions of the basic anomaly detection model and its metadata, and is responsible for receiving model fine-tuning feedback data from edge computing nodes in various regions for periodic optimization and version iteration of the basic model; The strategy management module is configured to define and distribute system operation strategies to edge computing nodes in each region. These strategies include a first preset threshold, a second preset threshold, data anonymization rules, and model update trigger conditions. The global monitoring module is configured to visually display the operating status, alarm statistics, and system performance indicators of edge computing nodes in each region, and provides historical data query and analysis functions.

[0013] Furthermore, the system operates on a hierarchical collaborative decision-making framework. This framework includes a real-time perception and preliminary decision-making layer located at camera nodes, a regional aggregation analysis and model fine-tuning layer located at regional edge computing nodes, and a global strategy management and model optimization layer located on the cloud management platform. The real-time perception and initial decision-making layer operates on a millisecond to second timescale, responsible for immediate anomaly detection and initial alarms. The regional aggregation analysis and model fine-tuning layer operates on a minute to hour timescale, responsible for cross-node correlation analysis and model personalization adaptation. The global strategy management and model optimization layer operates on a day to week timescale, responsible for system strategy formulation and basic model evolution.

[0014] Furthermore, the incremental learning training process adopted by the model fine-tuning module is as follows: an elastic weight consolidation algorithm is used to introduce regularization constraints on changes in important parameters when updating the basic anomaly detection model parameters using the local training dataset, in order to prevent catastrophic forgetting of previously learned knowledge and ensure that the fine-tuned model retains its original generalization ability while adapting to new local patterns.

[0015] Furthermore, the spatiotemporal correlation analysis process executed by the regional decision-making model is as follows: First, the edge decision data packets from different camera nodes are time-aligned and spatially mapped to construct a spatiotemporal event graph; then, a graph neural network is used to embed the spatiotemporal event graph to capture the potential propagation and correlation patterns between events; finally, the regional comprehensive risk index is calculated based on the output node representation of the graph neural network.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention deploys a local lightweight inference module on each camera node and performs data desensitization and feature extraction, thus pushing the processing of raw video data to the edge. Only desensitized feature data or lightweight decision results are uploaded, which greatly reduces the amount of data transmitted to the cloud, effectively alleviates network bandwidth pressure, reduces system operation and maintenance costs, and fundamentally cuts off the exposure path of raw privacy data on the transmission link, significantly enhancing user privacy protection.

[0017] 2. This invention provides each camera node with personalized model adaptation capabilities based on its own historical data through the model fine-tuning module in the regional edge computing node. This enables the anomaly detection model to continuously learn and adapt to the unique attributes of its deployment environment, such as specific lighting conditions, viewpoint changes and common activity patterns. This significantly improves the detection accuracy and robustness of the model in specific scenarios and overcomes the shortcomings of centralized models in terms of generalization ability.

[0018] 3. The hierarchical collaborative decision-making framework and incremental learning mechanism adopted in this invention realize multi-timescale collaborative operation from real-time perception to regional aggregation and then to global optimization. This not only ensures rapid local response to abnormal events, but also improves the identification accuracy of complex events through regional correlation analysis. At the same time, the continuous model iteration in the cloud ensures the overall evolution capability of the system, thus building an efficient, adaptive and sustainably optimized intelligent monitoring ecosystem. Attached Figure Description

[0019] Figure 1 This is a block diagram of the overall system architecture of the present invention; Figure 2 This is a diagram of the hierarchical collaborative decision-making framework of the present invention; Figure 3 This is a block diagram of the internal structure of the camera node in this invention; Figure 4 This is a block diagram of the internal structure of the edge computing node in the region of this invention; Figure 5 This is a block diagram of the internal structure of the cloud management platform of this invention. Detailed Implementation

[0020] Please refer to the attached document. Figures 1 to 5This embodiment details the specific technical implementation of an IoT-based intelligent camera monitoring system. The system aims to construct a distributed intelligent processing architecture, effectively addressing the problems faced by traditional centralized monitoring systems, such as massive data transmission pressure, user privacy leakage risks, and insufficient model generalization capabilities, by offloading data processing and analysis tasks to the network edge.

[0021] The system mainly consists of multiple camera nodes deployed in the monitored area, regional edge computing nodes communicating with these camera nodes, and a cloud management platform communicating with the regional edge computing nodes. These three core components form the system's perception layer, edge processing layer, and central coordination layer, respectively, working together to complete the entire process from data acquisition to intelligent decision-making.

[0022] Please refer to the attached document. Figure 3 As the perception layer unit of the system, camera nodes are directly deployed in various monitoring scenarios, responsible for the acquisition and preliminary intelligent processing of raw video data. Each camera node integrates a video acquisition module, a first preprocessing module, and a local lightweight inference module.

[0023] The video acquisition module typically employs a combination of a high-resolution image sensor and an optical lens, configured to continuously capture raw video streams of the monitored scene at a preset frame rate. This preset frame rate can be dynamically adjusted between 1 frame per second and 30 frames per second, with a typical value of 15 frames per second, to balance smoothness and data processing burden. The video acquisition module transmits the captured raw video stream data to the first preprocessing module in real time via an internal bus.

[0024] The first preprocessing module, connected to the video acquisition module, is a key component for data simplification and privacy protection at the camera nodes. This module is configured to perform two core operations on the received raw video stream: data anonymization and feature extraction. The data anonymization operation specifically includes real-time identification and blurring of sensitive information in the video stream.

[0025] The system utilizes pre-loaded face detection and license plate recognition models to scan video data frame by frame, accurately locating face and license plate regions. Once these sensitive areas are identified, Gaussian blur or pixelation algorithms are immediately applied to process them, generating de-identified video data that cannot reveal personal information. This process is completed locally, ensuring that the original privacy data never leaves the camera node.

[0026] The feature extraction operation follows the data anonymization operation and is configured to extract representative key information from the anonymized video data. Specifically, the first preprocessing module embeds a pre-built primary feature extraction network, which typically employs a lightweight convolutional neural network architecture, such as a cropped variant of MobileNet or SqueezeNet. This network receives anonymized video frames as input and outputs two main results: first, keyframes selected from the continuous video stream, representing significant changes or important events in the scene; and second, a multi-dimensional feature vector corresponding to each keyframe.

[0027] This multidimensional feature vector is a high-dimensional numerical array, typically ranging from 128 to 1024 dimensions, containing three main categories of features: scene semantic features, used to describe the object categories and scene context in the frame; motion optical flow features, used to characterize the motion patterns and velocities of objects between adjacent frames; and temporal context features, used to encode the positional relationships and short-term historical information of frames in the time series. The first preprocessing module packages the extracted keyframes and their corresponding multidimensional feature vectors and sends them to the local lightweight inference module via an internal interface.

[0028] The local lightweight inference module, connected to the first preprocessing module, is the core of the camera node's real-time intelligent analysis. This module is configured to load and run a lightweight anomaly detection model. This model is typically a simplified binary classifier or regressor with its parameter count strictly controlled to accommodate the limited computing resources of the camera node. The local lightweight inference module receives a multi-dimensional feature vector from the first preprocessing module as model input. The model calculates the input features and outputs a preliminary anomaly probability score between 0 and 1, quantifying the likelihood of abnormal behavior in the current monitored scene.

[0029] The local lightweight inference module has a pre-set first threshold, which is uniformly configured and distributed by the cloud management platform, typically ranging from 0.5 to 0.8. The module continuously compares the calculated preliminary anomaly probability score with this first preset threshold. When the preliminary anomaly probability score does not exceed the first preset threshold, the system determines that the current scenario is normal, and the relevant data is only briefly cached within the node or directly discarded.

[0030] Once the initial anomaly probability score exceeds a first preset threshold, the local lightweight inference module immediately triggers a two-stage response: First, it generates an audible and visual alarm signal locally to alert on-site personnel; second, it generates a structured edge decision data packet. This data packet contains at least the keyframe image that triggered the alarm, the corresponding multi-dimensional feature vector, the calculated initial anomaly probability score, the camera node's unique identifier, and a timestamp.

[0031] Subsequently, the local lightweight inference module uploads the edge decision data packets to the edge computing node of its logically belonging region via its integrated network communication unit. The entire process from data acquisition to decision upload is required to be completed within milliseconds to seconds to ensure rapid response to abnormal events.

[0032] Please refer to the attached document. Figure 4 With appendix Figure 5 Regional edge computing nodes constitute the edge processing layer of the system. They are typically deployed in network equipment rooms or base stations close to the camera node cluster and possess stronger computing and storage capabilities than a single camera node. Each regional edge computing node is responsible for managing multiple camera nodes within a geographical area or logical group, and internally includes a data aggregation module, a model fine-tuning module, and a second decision module.

[0033] The data aggregation module, serving as the data entry point for the regional edge computing nodes, is configured to continuously receive and cache edge decision data packets uploaded from all camera nodes under its jurisdiction. This module maintains a first-in, first-out (FIFO) data queue and allocates independent storage space for each camera node to temporarily store these data packets.

[0034] The data aggregation module performs preliminary verification on the received data packets, checking their format integrity and source legitimacy to ensure the reliability of subsequent processing. After successful verification, the data aggregation module forwards the edge decision data packets to the model fine-tuning module for model updates, and simultaneously forwards them to the second decision module for aggregation analysis.

[0035] The model fine-tuning module is the core engine for enabling personalized adaptation of camera node models. This module is configured to periodically perform a personalized model adaptation process for each camera node it manages. The core objective of this process is to fine-tune the general-purpose basic anomaly detection model using data generated by the individual camera node itself, making it more suitable for the unique environment in which that node operates.

[0036] The specific implementation steps are as follows: First, the model fine-tuning module periodically communicates with the cloud management platform to obtain the latest version of the basic anomaly detection model from the model repository module of the cloud management platform. This basic model is a relatively complex model, which has stronger feature representation capabilities and higher initial accuracy compared to the lightweight model running on the camera nodes.

[0037] Subsequently, the model fine-tuning module constructs a local training dataset for the camera node. This dataset consists of two parts: first, multi-dimensional feature vectors extracted from the data aggregation module, which are contained in the edge decision data packets generated by the camera node in recent times, such as the past 7 days; and second, the manually verified labels corresponding to these feature vectors.

[0038] The manually verified tags originate from alarm results confirmed by the administrator and fed back from the global monitoring module of the cloud management platform, or from genuine tags obtained through other trusted channels. These tags indicate whether the scenario represented by the corresponding feature vector is normal or abnormal.

[0039] Next, the model fine-tuning module uses this local training dataset to incrementally train the acquired basic anomaly detection model. To prevent the model from forgetting previously mastered common anomaly patterns when learning new knowledge, this system employs an elastic weight consolidation algorithm as the core mechanism for incremental learning. During training, this algorithm identifies parameters in the model that are more important to previous tasks and imposes constraints on the changes of these important parameters during updates.

[0040] Let the parameters of the basic model be... The local training dataset is The loss function is The elastic weight consolidation algorithm minimizes the loss on new data with each parameter update. At the same time, a regularization term is introduced to penalize important parameters for deviating from their initial values. The degree of. Its objective function can be expressed as: ; in, It is a hyperparameter that weighs the importance of new and old tasks. These are the diagonal elements of the Fisher information matrix, used to measure parameters. The importance of this approach to the old task is highlighted. By solving this optimization problem, the model can adapt to the local data distribution specific to each camera node while retaining as much general knowledge as possible from the base model. After training, the model fine-tuning module generates a set of fine-tuned model parameters specific to that camera node.

[0041] Finally, the module sends these fine-tuned model parameters to the corresponding camera nodes via a secure channel. Upon receiving the new parameters, the local lightweight inference module of the camera node will update the model during idle periods or as instructed, thereby improving the accuracy of its subsequent anomaly detection.

[0042] The second decision-making module is configured in the regional edge computing node and is responsible for performing cross-camera aggregation analysis and final decision-making. Its analysis complexity and field of view are higher than those of a single camera node. This module receives all edge decision data packets forwarded by the data aggregation module and runs a regional-level decision model. This regional-level decision model is typically a more complex machine learning model or rule engine that can comprehensively analyze data from multiple camera nodes within the same region.

[0043] The core task of the regional decision-making model is to perform spatiotemporal correlation analysis and calculate a comprehensive regional risk index. The spatiotemporal correlation analysis process is as follows: First, edge decision data packets from different camera nodes are time-aligned to unify all events onto the same time base, and spatial coordinate mapping is performed to locate the events detected by each camera node in a unified regional map coordinate system. Based on the aligned temporal and spatial information, a spatiotemporal event graph is constructed. In the graph, nodes represent abnormal events reported by each camera node, and edges represent the potential temporal and spatial correlations between events.

[0044] Subsequently, a graph neural network is used to perform embedding learning on the constructed spatiotemporal event graph. The graph neural network aggregates neighbor information for each node through a multi-layer message passing mechanism, thereby capturing complex propagation and correlation patterns between events. For example, it might discover a strong correlation between a person running event reported by camera node A and a vehicle abnormally stopping event reported by camera node B, after a specific time delay.

[0045] Finally, based on the final representation vector of each node output by the graph neural network, a scalar value, namely the regional comprehensive risk index, is calculated through a fully connected layer or aggregation function. This index comprehensively reflects the overall security situation of the entire monitored area.

[0046] The second decision-making module has a preset second threshold, which is also defined by the cloud management platform's policy management module. The module continuously compares the calculated regional comprehensive risk index with this second preset threshold. When the index does not exceed the threshold, the system considers the overall regional risk to be controllable. Once the regional comprehensive risk index exceeds the second preset threshold, the second decision-making module immediately generates a high-level alarm signal.

[0047] This signal not only includes the risk index itself, but also provides handling suggestions. These suggestions may include key patrol areas, a list of camera nodes that need to be monitored in conjunction with other systems, and inferences about potential event types. Finally, the second decision-making module uploads this high-level alarm signal, handling suggestions, and related data summaries to the cloud management platform via the uplink network interface of the regional edge computing node.

[0048] Please refer to the attached document. Figure 1 With appendix Figure 5 The cloud management platform forms the central coordination layer of the system. It is typically deployed in a cloud data center, possessing powerful computing and storage resources, and is responsible for the overall management and optimization of the system. The cloud management platform includes a model repository module, a policy management module, and a global monitoring module.

[0049] The model repository module serves as the management center for the system's foundational models. It is configured to store multiple versions of the basic anomaly detection models and their metadata. The metadata includes the model's version number, training data description, performance metrics, and applicable scenario descriptions. This module also receives model fine-tuning feedback data from edge computing nodes in various regions. This data includes loss curves and parameter change statistics generated during the fine-tuning process.

[0050] The cloud management platform periodically, for example every two weeks, aggregates feedback data from all regions and uses this data to retrain or fine-tune the basic anomaly detection model, generating a new version of the model with better performance and completing the iterative update of the basic model.

[0051] The policy management module is the center for formulating and distributing system operation rules. This module is configured to define and issue a series of system operation policies to all regional edge computing nodes, which in turn further distribute them to their subordinate camera nodes.

[0052] These strategies include, but are not limited to: the first preset threshold used by the local lightweight inference module of the camera node, the second preset threshold used by the second decision module of the regional edge computing node, the data anonymization rules executed by the first preprocessing module, and the conditions for the model fine-tuning module to trigger the model update task. The strategy management module provides a graphical interface for administrators to configure these parameters and ensures that strategy changes can be securely and consistently pushed to all nodes in the network.

[0053] The global monitoring module serves as the system's status visualization and data analysis center. This module is configured to receive real-time heartbeat information, status reports, and high-level alarm signals from edge computing nodes across various regions. It visualizes the operational status of all edge computing nodes in the entire system through a centralized dashboard interface, including online / offline status, CPU and memory utilization. Alarm statistics, such as the number, distribution, and trends of various alarms, as well as overall system performance indicators. This module also provides powerful historical data query and analysis functions, allowing administrators to retrospectively analyze historical events by time, region, alarm type, and other dimensions, providing data support for optimizing monitoring strategies.

[0054] Please refer to the attached document. Figure 2 The entire system operates within a rigorous hierarchical collaborative decision-making framework. This framework clearly defines three functional and time scales. The real-time perception and preliminary decision-making layer, located at the camera nodes, operates on a millisecond to second time scale. Its core task is to achieve immediate anomaly detection and initial alarm, ensuring a rapid response to emergencies.

[0055] The regional aggregation analysis and model fine-tuning layer, located at the edge computing nodes, operates on a timescale of minutes to hours. It is responsible for cross-camera node correlation analysis to improve the accuracy of complex event recognition and for performing model personalization adaptation to optimize the detection performance of individual nodes. The global policy management and model optimization layer, located on the cloud management platform, operates on a timescale of days to weeks. It focuses on system-level policy formulation, continuous evolution of the basic model, and monitoring and analysis of the global status.

[0056] These three levels each have their own responsibilities and work closely together to form an efficient, adaptive, and continuously optimized intelligent monitoring ecosystem.

[0057] This embodiment provides another specific implementation plan for an IoT-based intelligent camera monitoring system. Based on the above embodiment, it mainly optimizes and refines the spatiotemporal correlation analysis process of the regional-level decision model in the regional edge computing node and the incremental learning strategy of the model fine-tuning module to adapt to the needs of more complex city-level large-scale monitoring networks.

[0058] In the second decision module of the regional edge computing node, the spatiotemporal correlation analysis process performed by the regional decision model is further enhanced. In addition to constructing a basic spatiotemporal event graph and performing graph neural network embedding learning, this embodiment introduces a multimodal feature fusion mechanism.

[0059] Specifically, when constructing the spatiotemporal event graph, the node attributes are no longer limited to preliminary anomaly probability scores, but incorporate deeper information from multidimensional feature vectors in the edge decision data package. The system utilizes a pre-trained encoder to map multidimensional feature vectors into more semantically informative initial node features.

[0060] Meanwhile, the weight calculation of edges in the graph is more refined, considering not only Euclidean spatial distance and temporal difference between events, but also introducing scene semantic similarity as a weighting factor. For example, two events reporting "crowd gathering" and "vehicle loitering" respectively, even if they are close in time and space, may have a lower association weight than two events both reporting "rapid movement." This allows the constructed spatiotemporal event graph to more accurately reflect the complex relationships between real-world events.

[0061] The structure of the graph neural network has also been adjusted accordingly, adopting a graph convolutional network enhanced with an attention mechanism. During message passing, nodes assign different attention weights based on the importance of the features of their neighboring nodes, thus focusing more on the information most relevant to the current risk assessment. The calculation of the regional comprehensive risk index is no longer a simple linear aggregation, but introduces a small recurrent neural network layer to model the dynamic evolution trend of the regional risk status in a short period of time, making the risk index more forward-looking.

[0062] Regarding the model fine-tuning module, this embodiment optimizes the incremental learning training process to address the potential non-stationary distribution of camera node data streams. In addition to continuing to employ the elastic weight consolidation algorithm to overcome catastrophic forgetting, the system also integrates a dynamic learning rate adjustment strategy and a simple rollback mechanism. The dynamic learning rate adjustment strategy adaptively adjusts the learning rate during training based on the size of the local training dataset, data distribution, and historical fine-tuning results, avoiding overfitting when data volume is low or noise levels are high.

[0063] The rollback mechanism automatically reverts the update and restores the previous version's model parameters when a significant performance degradation is detected on a retained validation set after the current fine-tuning. It also records this anomaly and reports it to the cloud management platform for analysis and processing. This enhances the robustness of the personalized model adaptation process.

[0064] Furthermore, this embodiment extends the functionality of the policy management module of the cloud management platform. In addition to issuing static thresholds and rules, the policy management module also supports issuing lightweight, scenario-specific rule templates to regional edge computing nodes. Regional edge computing nodes can then allocate rules based on the characteristics of their managed areas. For example, templates can be instantiated for commercial areas, residential areas, or transportation hubs to form localized analysis rules that better fit the characteristics of the region. These rules can work in conjunction with regional decision-making models. For instance, rules can be defined to directly trigger high-level alarms when a specific combination of certain types of events occurs in a certain area within a specific time period, thereby improving the system's response speed to known typical threat patterns.

[0065] At the system communication level, this embodiment strengthens the data compression and encryption mechanisms between different layers. Before encapsulation, the edge decision data packets uploaded by the camera nodes to the regional edge computing nodes employ a more efficient compression algorithm on the key frame images and perform end-to-end encryption on the entire data packet.

[0066] The communication between the regional edge computing nodes and the cloud management platform also adopts a high-strength encryption protocol, and further aggregates and samples the uploaded data digests to minimize the amount of uplink data while ensuring that key information is not lost, so as to adapt to the network bottlenecks that may occur under large-scale deployment.

[0067] Finally, the global monitoring module in this embodiment enhances its data analysis capabilities by integrating root cause analysis functionality. When a high-level alarm occurs, this function can automatically trace all underlying camera node events related to the alarm, the model's decision-making basis, and the policy triggering conditions, displaying the event chain graphically. This system assists administrators in quickly understanding the root causes of alarms, assessing the rationality of system decisions, and providing precise guidance for further optimization of system parameters and models. These enhancements enable the system described in this embodiment to possess higher levels of intelligence, adaptability, and operational efficiency in large-scale, heterogeneous city-level surveillance applications.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A camera-based intelligent monitoring system based on the Internet of Things, characterized in that, The system includes multiple camera nodes deployed in the monitoring area, at least one regional edge computing node communicatively connected to the camera nodes, and a cloud management platform communicatively connected to the regional edge computing node; the camera nodes constitute the perception layer of the system, the regional edge computing node constitutes the edge processing layer of the system, and the cloud management platform constitutes the central coordination layer of the system; each camera node integrates a video acquisition module, a first preprocessing module, and a local lightweight inference module. The video acquisition module is configured to continuously capture the original video stream of the monitored scene at a preset frame rate; the first preprocessing module is connected to the video acquisition module and is configured to perform data desensitization and feature extraction operations on the original video stream. The data desensitization operation includes real-time blurring of the face and license plate regions in the video stream to generate desensitized video data; the feature extraction operation includes using a pre-set primary feature extraction network to extract keyframes and their corresponding multi-dimensional feature vectors from the desensitized video data, wherein the multi-dimensional feature vectors include scene semantic features, motion optical flow features and temporal context features. The local lightweight inference module is connected to the first preprocessing module and is configured to load and run a lightweight anomaly detection model. The lightweight anomaly detection model receives the multi-dimensional feature vector output by the first preprocessing module as input and outputs a preliminary anomaly probability score. When the preliminary anomaly probability score exceeds a first preset threshold, the local lightweight inference module triggers an alarm and generates an edge decision data packet containing keyframes, multi-dimensional feature vectors, and a preliminary anomaly probability score, and simultaneously uploads the edge decision data packet to its corresponding regional edge computing node.

2. The IoT-based intelligent camera monitoring system according to claim 1, characterized in that, It also includes local lightweight inference models and regional decision-making models; The regional edge computing node includes a data aggregation module, a model fine-tuning module, and a second decision module; The data aggregation module is configured to receive and cache edge decision data packets from all camera nodes under its jurisdiction. The model fine-tuning module is configured to perform a personalized model adaptation process for each camera node; The second decision module is configured to perform aggregate analysis and final decision.

3. The IoT-based intelligent camera monitoring system according to claim 2, characterized in that, The personalized model adaptation process is as follows: The model fine-tuning module obtains the basic anomaly detection model from the cloud management platform, and uses the historical feature vectors in the edge decision data packets recently generated by the camera node and the manual review labels to form a local training dataset. It performs incremental learning training on the basic anomaly detection model, generates fine-tuning model parameters specific to the camera node, and sends the fine-tuning model parameters to the local lightweight inference module of the corresponding camera node for model update.

4. The IoT-based intelligent camera monitoring system according to claim 3, characterized in that, The incremental learning training process employs an elastic weight consolidation algorithm. When updating the parameters of the basic anomaly detection model using the local training dataset, regularization constraints on changes in important parameters are introduced to prevent catastrophic forgetting of previously learned knowledge.

5. The IoT-based intelligent camera monitoring system according to claim 2, characterized in that, The second decision module receives edge decision data packets forwarded by the data aggregation module and runs a regional decision model with a higher complexity than the local lightweight inference model; the regional decision model comprehensively analyzes edge decision data packets from multiple camera nodes in the same area, performs spatiotemporal correlation analysis, and calculates the regional comprehensive risk index. The second decision module has a second preset threshold. When the regional comprehensive risk index exceeds the second preset threshold, the second decision module generates a high-level alarm signal and a handling suggestion, and uploads the high-level alarm signal, the handling suggestion and related data summary to the cloud management platform.

6. The IoT-based intelligent camera monitoring system according to claim 5, characterized in that, The spatiotemporal correlation analysis process is as follows: First, the edge decision data packets from different camera nodes are time-aligned and spatially mapped to construct a spatiotemporal event graph. Subsequently, a graph neural network is used to embed the spatiotemporal event graph to capture the potential propagation and correlation patterns between events; finally, the comprehensive risk index of the region is calculated based on the output node representation of the graph neural network.

7. The IoT-based intelligent camera monitoring system according to claim 5, characterized in that, The cloud management platform includes a model repository module, a strategy management module, and a global monitoring module. The model repository module stores multiple versions of the basic anomaly detection model and its metadata, and is responsible for receiving model fine-tuning feedback data from edge computing nodes in various regions for the periodic optimization and version iteration of the basic model. The policy management module is configured to define and distribute system operation policies to edge computing nodes in each region. The global monitoring module is configured to visually display the operating status, alarm statistics and system performance indicators of edge computing nodes in each region, and provide historical data query and analysis functions.

8. The IoT-based intelligent camera monitoring system according to claim 7, characterized in that, The system operation strategy includes the first preset threshold, the second preset threshold, data anonymization rules, and model update triggering conditions.

9. The IoT-based intelligent camera monitoring system according to claim 1, characterized in that, The system operates on a hierarchical collaborative decision-making framework, which includes a real-time perception and preliminary decision-making layer located at camera nodes, a regional aggregation analysis and model fine-tuning layer located at regional edge computing nodes, and a global strategy management and model optimization layer located on the cloud management platform. The real-time perception and preliminary decision-making layer operates on a timescale of milliseconds to seconds; the regional aggregation analysis and model fine-tuning layer operates on a timescale of minutes to hours; and the global strategy management and model optimization layer operates on a timescale of days to weeks.

10. A camera-based intelligent monitoring method based on the Internet of Things, characterized in that, The camera intelligent monitoring system based on the Internet of Things as described in any one of claims 1-9 is used to achieve intelligent camera monitoring.

Images