A police video image intelligent analysis system based on artificial intelligence

Abstract

The application relates to the technical field of video image analysis, and provides a police video image intelligent analysis system based on artificial intelligence, which realizes real-time sensing and early warning of police situations by performing target detection, identity recognition and abnormal behavior analysis on video image data on an edge computing node. In the target detection, a physical consistency constraint and a continuous knowledge distillation mechanism are introduced to improve the detection stability of the target in the continuous time and space dimensions; in the identity recognition, cross-layer tensor orthogonalization optimization is used to enhance the distinctness and consistency of the identity feature representation. In combination with the edge computing deployment and the blockchain storage mechanism, low-latency response and data credible management of police situation analysis are realized, and the system is suitable for complex police monitoring and public security application scenarios.

Description

Technical Field

[0001] This invention relates to the field of video image analysis technology, and in particular to an intelligent analysis system for police video images based on artificial intelligence. Background Technology

[0002] With the continuous increase in the demand for public safety governance, video surveillance systems have been widely deployed in public places, major traffic arteries, and law enforcement sites. Traditional video surveillance technology mainly relies on manual patrols or rule-based image processing methods to analyze video content, typically only providing post-event retrieval and playback functions, making it difficult to cope with the real-time analysis needs of large-scale, long-term video data. In recent years, artificial intelligence technology has been gradually introduced into the field of video surveillance. Deep learning-based object detection, image recognition, and behavior analysis methods have achieved a certain degree of automated identification of people, vehicles, and objects, showing significant improvements in detection accuracy and automation compared to traditional methods. However, existing intelligent video analysis technologies mostly focus on modeling the features of single frames or local time windows, lacking unified constraints on the evolution of targets in continuous time and space. In complex dynamic monitoring scenarios, this can easily lead to problems such as inconsistent detection results across frames, target jitter, short-term loss, and unstable categories, thus affecting the reliability of identity recognition and behavior analysis. At the same time, the complex model structures and high-order constraint mechanisms introduced to improve detection accuracy often result in increased computational overhead and inference latency, making it difficult to directly adapt to edge computing deployment scenarios with high real-time requirements. In addition, existing intelligent systems mostly adopt centralized storage for the data management of police situation analysis results, lacking anti-tampering evidence preservation and consistency verification mechanisms for key analysis results, which makes it difficult to meet the requirements of law enforcement scenarios for data credibility, security, and compliance. Therefore, based on a comparison of traditional video surveillance technology and existing intelligent video analysis technology, there is still a need for a police video image intelligent analysis technology solution that can balance detection stability, real-time performance, and data credibility in complex environments. Summary of the Invention

[0003] This invention addresses the problems of inconsistent target detection results across frames, difficulty in balancing model inference efficiency and stability, and insufficient reliability of police incident analysis data in existing intelligent analysis technologies for police video images under complex dynamic monitoring scenarios. It proposes an AI-based intelligent analysis system for police video images, achieving a unified improvement in the stability, real-time performance, and reliability of intelligent video analysis through collaborative design at the model training and system architecture levels. Regarding the intelligent analysis model, a training mechanism combining physical consistency constraints and continuous knowledge distillation is introduced. The evolutionary patterns of targets in continuous time and space dimensions are incorporated into the teacher model training process, and spatiotemporal consistency features are transferred to the lightweight student model through a continuous consistency representation space. This allows for stable target detection results across frames without the need for high-order physical constraint calculations during the inference stage. While ensuring detection consistency, the system reduces computational complexity and inference latency. In identity recognition, a cross-layer tensor orthogonalization optimization mechanism is introduced during the training of the visual Transformer coding network. This mechanism imposes cross-layer joint constraints on the parameter update direction of the isomorphic coding layer, suppressing redundancy and overlap in deep feature learning and enhancing the discriminability and stability of identity feature representation under complex lighting, pose changes, and occlusion conditions. At the system application level, by deploying target detection, identity recognition, and abnormal behavior analysis on edge computing nodes, the system achieves rapid perception and early warning of abnormal behavior. Furthermore, blockchain technology is used to encrypt and verify the results of the police incident analysis and its associated data, thereby improving the data credibility, security, and compliance of the police video intelligent analysis system in law enforcement applications while meeting real-time response requirements.

[0004] This invention provides an intelligent analysis system for police video images based on artificial intelligence. The system includes: a video acquisition module, an edge computing module, an alarm warning module, a monitoring center, and terminal equipment. The video acquisition module collects real-time video image data through high-definition camera equipment deployed in public places, traffic arteries, law enforcement sites, and monitoring areas; The edge computing module includes a high-performance processor, a storage module, and an interface unit for connecting to external devices. It receives video image data and performs preprocessing and feature extraction on the video image data locally to obtain preprocessed feature data. The preprocessed feature data includes spatial features, motion features, color histogram features, edge features, texture features, and illumination features extracted from the video image data, which are used for subsequent target detection and behavior analysis tasks. The edge computing module also includes an artificial intelligence analysis module, which includes a target detection unit, an image recognition unit, and an abnormal behavior analysis unit. The target detection unit employs a consistent target detection CNN model to perform temporal consistency analysis on the features of consecutive frames in the preprocessed feature data, performs region localization processing on the target, determines the corresponding target region image, and outputs the target detection result corresponding to the target region image. The target detection result includes the target's location coordinates, target category identifier, and target confidence information. The targets include personnel targets, vehicle targets, and portable object targets. The consistent target detection CNN model includes the Teacher-CNN-PINN model and the Student-CNN model. The image recognition unit constructs a visual Transformer encoding network and introduces a cross-layer tensor orthogonalization optimization method during the training process of this encoding network. This method performs cross-layer joint constraints and orthogonalization optimization on the parameter update process within the isomorphic encoding layers of the visual Transformer encoding network, resulting in a trained visual Transformer encoding network. The trained visual Transformer encoding network processes the target region image, generating a fixed-dimensional feature vector representation. A metric learning mechanism is used to calculate the similarity of the fixed-dimensional feature vector representation and match it with the historical feature vector library maintained in the storage module. This completes the target category attribute confirmation and target identity feature consistency determination, outputting the identity recognition result corresponding to the target region image. The identity recognition result includes target identity feature description information, target category confirmation result, and feature matching similarity result. The visual Transformer encoding network encodes the target's appearance components, texture structure, and contextual associations through a multi-head self-attention mechanism. The abnormal behavior analysis unit, based on the target detection results and identity recognition results, models and analyzes the spatial position changes and behavioral trajectories of the target within a continuous time period, identifies whether the target behavior meets the preset abnormal behavior judgment conditions, and outputs the abnormal behavior analysis results, including abnormal behavior type identifier, occurrence time information and corresponding video image segment index information; The alarm warning module is linked to the alarm system and database. Based on the abnormal behavior analysis results output by the edge computing module, it automatically triggers alarms and generates alarm reports. The alarm reports are then transmitted to the monitoring center and terminal device modules to achieve real-time response. The command and decision-making interaction module, deployed in the monitoring center and working in conjunction with terminal devices, is used to perform predictive analysis and risk assessment of emergencies based on historical alarm reports and abnormal behavior analysis results, combined with machine learning methods, to generate auxiliary decision-making information. The machine learning methods include a prediction model based on time-series feature modeling and a risk assessment model based on supervised learning, used to predict trends and assess risk levels based on historical alarm reports and abnormal behavior analysis results. This module displays real-time video images, alarm reports, historical data, and analysis results to relevant personnel through a visual monitoring platform, and combines auxiliary decision-making information for manual intervention and command transmission to support rapid response and on-site handling. The terminal devices include handheld terminals, vehicle-mounted terminals, and monitoring display terminals. The handheld terminals are smartphones and tablets, equipped with integrated real-time video transmission and interaction functions, supporting remote viewing of video images, receiving alarm alerts, issuing commands, and video image playback, meeting the needs of mobile office and remote command. The data transmission module is used for data communication between the video acquisition module, edge computing module, alarm warning module, and command and decision interaction module. This module includes an Ethernet wired transmission unit and a fifth-generation mobile communication wireless transmission unit, which work together during system operation. The system uses blockchain technology to encrypt and store the police reports generated by the police alert module and their associated target detection results, identity recognition results and abnormal behavior analysis results, and verify their consistency. This ensures the immutability, trustworthiness and compliance of the data, and meets the data privacy and security requirements in the law enforcement process.

[0005] Furthermore, the process of using a consistent target detection CNN model to perform temporal consistency analysis on the preprocessed feature data for continuous frame features, performing target region localization processing, determining the corresponding target region image, and outputting the target detection result corresponding to the target region image specifically includes the following steps: Step S1: Perform time synchronization, spatial registration and scale normalization on the preprocessed feature data to establish a unified spatiotemporal reference system and form standardized feature data; Step S2: Using the CNN model in the Teacher-CNN-PINN model as the backbone network for image feature extraction, perform multi-layer convolution, feature mapping, and target region encoding on the standardized feature data to generate a high-dimensional feature representation for target localization and category discrimination; construct a PINN physical consistency constraint module in the output layer of the CNN model to formalize the evolution law of the target region in continuous time and space dimensions into a continuously differentiable physical residual operator, introduce the target detection and recognition loss term and combine it with the physical residual operator to construct a joint target loss function, which is used as the training objective function of the Teacher-CNN-PINN model and participates in backpropagation optimization to obtain the trained Teacher-CNN-PINN model; Step S3: Define a continuous consistency representation space. Using the trained Teacher-CNN-PINN model, perform forward inference to map the model's output at different times and spatial locations to the continuous consistency representation space, generating continuous consistency representation features for the corresponding target region. The continuous consistency representation features are the soft physical representation results output by the Teacher-CNN-PINN model, used to characterize the evolutionary relationship and consistency features of the target region in continuous time and spatial dimensions. Step S4: By performing soft-supervised learning on the continuous consistency representation features output by the Teacher-CNN-PINN model, the Student-CNN model implicitly inherits the spatiotemporal consistency features formed by the Teacher model under physical constraints without explicitly introducing a physical constraint module. Step S5: Based on the continuous and consistent representation space, a two-term coupling loss function is used to distill and optimize the Student-CNN model, so as to realize the transfer of the spatiotemporal consistent features learned in the Teacher-CNN-PINN model to the Student-CNN model. The two-term coupling loss function is constructed as follows: when distilling the Student-CNN model, the KD-PINN continuous distillation mechanism is used. Instead of introducing the discrete probability distribution form based on softmax, the continuous and consistent representation features output by the Teacher-CNN-PINN model at the corresponding spatiotemporal position are used as the soft supervision target of the Student-CNN model, so that the distillation process is carried out in a continuous and differentiable output space. Based on the consistency constraint between the Teacher-CNN-PINN model and the Student-CNN model in the continuous output space, a continuous distillation loss function between the Student-CNN model and the Teacher-CNN-PINN model is constructed. The continuous distillation loss function is jointly modeled with the target detection and recognition loss term of the Student-CNN model to construct the two-term coupling loss function. Step S6: After the distillation training is completed, the trained Student-CNN model is obtained. The trained Student-CNN model is used to perform forward inference processing, perform consistent target detection analysis on the target in the continuous frames, automatically locate the target region image corresponding to the target, and output the target detection result corresponding to the target region image.

[0006] Furthermore, in the image recognition unit, a cross-layer tensor orthogonalization optimization method is introduced to perform cross-layer joint constraints and orthogonalization optimization on the parameter update process within the isomorphic coding layer of the visual Transformer's coding network, resulting in the trained visual Transformer's coding network. The process of processing the target region image using the trained visual Transformer's coding network to generate a fixed-dimensional feature vector representation specifically includes the following steps: Step B1: In the visual Transformer encoding network, identify consecutive isomorphic Transformer encoding layers, and determine the linear projection weight matrix for multi-head self-attention calculation in each encoding layer. The set of linear projection weight matrices includes a query matrix, a key matrix, and a value matrix. Perform cross-layer grouping on the isomorphic Transformer encoding layers according to a preset group size to form a layer group index interval. Step B2: During the training phase, the target recognition loss function is introduced to perform backpropagation calculation on the encoding network of the visual Transformer. For each isomorphic encoding layer within the index interval of each layer group, the gradient information corresponding to its linear projection weight matrix is ​​extracted to form a set of gradient matrices divided by layer, which is used for subsequent cross-layer joint optimization processing. Step B3: Perform momentum accumulation processing on the gradient matrix set, and jointly organize the momentum matrices corresponding to each layer according to the layer index order. Introduce the cross-layer momentum tensor construction formula to construct a third-order momentum tensor for cross-layer joint orthogonalization processing. Step B4: Perform matrix expansion on the third-order momentum tensor along mode-1 to obtain a flattened matrix; apply the gradient orthogonalization operator to the flattened matrix and perform cross-layer joint orthogonalization to eliminate redundant overlap of momentum in the main singular direction of different coding layers to obtain an orthogonalized matrix; perform mode-1 inverse matrixization on the orthogonalized matrix to obtain an orthogonalized momentum tensor. Step B5: During the training phase, based on the orthogonalized momentum tensor, perform parameter update operations on each isomorphic Transformer encoding layer within the layer group index interval to complete the optimized training of the visual Transformer encoding network and obtain the trained visual Transformer encoding network. Step B6: During the inference phase, the encoding network of the trained visual Transformer is used to perform forward encoding processing on the target region image, outputting a global token feature vector and a local aggregated feature vector respectively; the global feature vector and the local feature vector are subjected to weighted fusion and normalization processing to generate a fixed-dimensional feature vector representation.

[0007] By adopting the above solution, the beneficial effects achieved by the present invention are as follows: This invention introduces a consistent target detection mechanism that combines physical consistency constraints with continuous knowledge distillation into a police video image intelligent analysis system. This mechanism enables stable detection and localization of targets in continuous time and space dimensions, improving the consistency and reliability of target detection results under complex dynamic monitoring environments across frames and time sequences. It solves the problems of detection jitter, target loss, and category instability that easily occur in existing intelligent video analysis technologies when targets such as people, vehicles, and portable items move rapidly, change scale, or are occluded. This enhances the basic ability of the police situation analysis system to continuously track and accurately judge abnormal events, thereby providing stable and reliable target input for subsequent identity recognition, behavior analysis, and police situation early warning.

[0008] In the identity recognition process, this invention introduces a cross-layer tensor orthogonalization optimization mechanism during the training of the visual Transformer coding network. This achieves cross-layer joint constraints on the parameter update direction of the isomorphic coding layer, improving the discriminability and stability of identity feature representation under complex lighting changes, pose changes, and local occlusion conditions. It solves the problems of feature redundancy, decreased discriminative ability, and insufficient cross-scene generalization ability in existing identity recognition models in deep structures. This enhances the reliability of the police video image intelligent analysis system in determining the consistency of target identity features and the identity matching results, thereby improving the practical application effect of the system in police situation verification, key personnel identification, and multi-scene linkage analysis.

[0009] This invention deploys target detection, identity recognition, and abnormal behavior analysis functions on edge computing nodes, enabling local intelligent processing and rapid response of police video image data. This improves the real-time performance and engineering availability of the system in actual law enforcement and emergency scenarios, and solves the problems of high transmission pressure and high response latency faced by traditional centralized intelligent analysis architectures under high-concurrency video access conditions. At the same time, by introducing blockchain encryption and consistency verification mechanisms into the police situation analysis results and their related data, the immutability and trustworthiness of police situation data in the generation, transmission, and storage processes are enhanced. Thus, while meeting the needs of rapid police situation handling, it provides reliable guarantees for the compliance, security, and traceability of the police video intelligent analysis system in law enforcement applications. Attached Figure Description

[0010] Figure 1This is a schematic diagram of the overall structure of an intelligent analysis system for police video images based on artificial intelligence, as proposed in this invention. Figure 2 This is a schematic diagram comparing the distribution of principal directions before and after orthogonalization of the cross-layer momentum tensor proposed in Example 5. Detailed Implementation

[0011] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Example 1, according to Figure 1 This invention provides an intelligent analysis system for police video images based on artificial intelligence. The system includes: a video acquisition module, an edge computing module, an alarm warning module, a monitoring center, and terminal equipment. The video acquisition module uses high-definition cameras deployed in public places, traffic arteries, law enforcement sites, and monitored areas to collect real-time video image data. The module features adaptive adjustment of imaging parameters, including focal length, exposure, and gain. These parameters are automatically adjusted by the camera's internal control unit based on the brightness distribution, contrast distribution, and motion intensity characteristics of the captured image, maintaining stable image quality under various conditions such as day / night cycles, direct sunlight, backlighting, fog, rain, and snow. The module includes a rotatable 360-degree panoramic camera with a motorized pan-tilt mechanism. This mechanism supports continuous horizontal and vertical angle adjustment, achieving full coverage video capture of the monitored area. The panoramic camera maintains image continuity and time synchronization during rotation, ensuring the video image data is suitable for subsequent AI analysis and processing. The camera system employs a sealed protective structure, meeting waterproof, dustproof, and weather-resistant requirements, enabling stable operation in high-temperature, low-temperature, high-humidity, and strong-wind environments. The edge computing module, including a high-performance processor, storage module, and interface unit for connecting to external devices, receives video image data and performs preprocessing and feature extraction locally to obtain preprocessed feature data. This preprocessed feature data includes spatial features, motion features, color histogram features, edge features, texture features, and illumination features extracted from the video image data, used for subsequent target detection and behavior analysis tasks. The edge computing module internally includes an artificial intelligence analysis module, comprising a target detection unit, an image recognition unit, and an abnormal behavior analysis unit. Spatial features describe the spatial distribution and shape characteristics of targets in the image; motion features include the target's trajectory and speed changes, primarily used for dynamic behavior analysis; color histogram features represent the color distribution in the image, helping to identify the target's color pattern; edge features extract the target's boundary information through edge detection algorithms, helping to clarify the target's shape; texture features describe the surface details of the image, often used to identify objects and specific scenes; and illumination features reflect changes in illumination in the image, aiding in target detection under different lighting conditions. The target detection unit employs a consistent target detection CNN model to perform temporal consistency analysis on the features of consecutive frames in the preprocessed feature data, performs region localization processing on the target, determines the corresponding target region image, and outputs the target detection result corresponding to the target region image. The target detection result includes the target's location coordinates, target category identifier, and target confidence information. The targets include personnel targets, vehicle targets, and portable object targets. The consistent target detection CNN model includes the Teacher-CNN-PINN model and the Student-CNN model. The image recognition unit constructs a visual Transformer encoding network and introduces a cross-layer tensor orthogonalization optimization method during the training process of this encoding network. This method performs cross-layer joint constraints and orthogonalization optimization on the parameter update process within the isomorphic encoding layers of the visual Transformer encoding network, resulting in a trained visual Transformer encoding network. The trained visual Transformer encoding network processes the target region image, generating a fixed-dimensional feature vector representation. A metric learning mechanism is used to calculate the similarity of the fixed-dimensional feature vector representation and match it with the historical feature vector library maintained in the storage module. This completes the target category attribute confirmation and target identity feature consistency determination, outputting the identity recognition result corresponding to the target region image. The identity recognition result includes target identity feature description information, target category confirmation result, and feature matching similarity result. The visual Transformer encoding network encodes the target's appearance components, texture structure, and contextual associations through a multi-head self-attention mechanism. The abnormal behavior analysis unit, based on the target detection results and identity recognition results, models and analyzes the spatial position changes and behavioral trajectories of the target within a continuous time period, identifies whether the target behavior meets the preset abnormal behavior judgment conditions, and outputs the abnormal behavior analysis results, including abnormal behavior type identifier, occurrence time information and corresponding video image segment index information; In this embodiment, specifically during the time period from 22:41:18 to 22:41:52, the edge computing module receives continuous video image data from the south entrance of Station A. After analysis by the consistency target detection CNN model, the following target detection results are output: Target ID: P_20240218_0047; Target category identifier: Personnel target; Target confidence information: 0.94; Continuous frame consistency score: 0.91; The target maintains a stable detection result in continuous frames, and the target area positioning does not change, satisfying the temporal consistency constraint.

[0013] The identity recognition results are as follows: Target identity feature description information: Feature vector dimension: 512, Feature stability score: 0.88; Target category confirmation result: General public personnel; Identity matching similarity result: Maximum similarity with the key personnel database: 0.32, Maximum similarity with the historical alarm-related personnel database: 0.29; The system determines that the target does not belong to the registered key personnel database, but the identity features remain consistent in the current time period.

[0014] Based on the target detection and identity recognition results, the abnormal behavior analysis unit modeled and analyzed the target's spatial location changes and behavioral trajectory during the time period from 22:41:18 to 22:44:06, obtaining the following results: Behavioral trajectory characteristics: multiple back-and-forth trips between the ticket area and the security checkpoint, with a significantly longer dwell time per unit time than the average for the same area, and the behavioral path deviating from the normal passage path; Abnormal behavior type identifier: Suspicious loitering behavior; Abnormal behavior occurrence time information: Start time: 22:41:35, Judgment time: 22:43:12; Corresponding video image segment index information: Video stream ID: CAM_ST_S_03, Segment time interval: 22:41:18—22:44:06; The abnormal behavior analysis results were marked as "Attention Required Level".

[0015] The alarm warning module is linked to the alarm system and database. Based on the abnormal behavior analysis results output by the edge computing module, it automatically triggers alarms and generates alarm reports. The alarm reports are then transmitted to the monitoring center and terminal device modules to achieve real-time response. After receiving the abnormal behavior analysis results, the alarm warning module, in conjunction with the alarm system and database, automatically generates an alarm report. The report includes: alarm number: ALERT_20240218_221; alarm type: abnormal behavior warning in public places; target ID involved: P_20240218_0047; description of abnormal behavior: the target loitered in the rail transit station hall area for a long time, and the behavior pattern was significantly different from the normal passage pattern; time interval of occurrence: 22:41:18—22:44:06; the alarm report is encrypted and stored in the blockchain ledger, and is simultaneously pushed to the monitoring center and terminal devices.

[0016] The command and decision-making interaction module, deployed in the monitoring center and working in conjunction with terminal devices, is used to perform predictive analysis and risk assessment of emergencies based on historical alarm reports and abnormal behavior analysis results, combined with machine learning methods, to generate auxiliary decision-making information. The machine learning methods include a prediction model based on time-series feature modeling and a risk assessment model based on supervised learning, used to predict trends and assess risk levels based on historical alarm reports and abnormal behavior analysis results. This module displays real-time video images, alarm reports, historical data, and analysis results to relevant personnel through a visual monitoring platform, and combines auxiliary decision-making information for manual intervention and command transmission to support rapid response and on-site handling. The terminal devices include handheld terminals, vehicle-mounted terminals, and monitoring display terminals. The handheld terminals are smartphones and tablets, equipped with integrated real-time video transmission and interaction functions, supporting remote viewing of video images, receiving alarm alerts, issuing commands, and video image playback, meeting the needs of mobile office and remote command. The data transmission module facilitates data communication between the video acquisition module, edge computing module, alarm warning module, and command and decision interaction module. This module includes an Ethernet wired transmission unit and a 5G wireless transmission unit, which work collaboratively during system operation. The Ethernet wired transmission unit, using a gigabit Ethernet interface and based on the TCP / IP protocol, undertakes the primary transmission of video image data in fixed deployment scenarios. The 5G wireless transmission unit, based on the 5G cellular communication protocol, undertakes the data transmission of video image data in mobile deployment scenarios and takes over the real-time transmission of video image data when the wired transmission link is abnormal or unavailable. Based on link status detection results, this module performs transmission scheduling between the Ethernet wired transmission unit and the 5G wireless transmission unit to ensure continuous and stable transmission of video image data and analysis results to the edge computing module and command and decision interaction module.

[0017] The system uses blockchain technology to encrypt and store the police reports generated by the police alert module and their associated target detection results, identity recognition results and abnormal behavior analysis results, and verify their consistency. This ensures the immutability, trustworthiness and compliance of the data, and meets the data privacy and security requirements in the law enforcement process.

[0018] Example 2 differs from Example 1 in that the target detection unit uses an R-CNN model to perform temporal consistency analysis on the features of consecutive frames using preprocessed feature data. In this example, the target detection unit specifically includes the following: using an R-CNN model to perform temporal consistency analysis on the features of consecutive frames using preprocessed feature data, performing region localization processing on the target, determining the corresponding target region image, and outputting the target detection result corresponding to the target region image. The target detection result includes the target's location coordinates, target category identifier, and target confidence information.

[0019] Example 3 differs from Example 1 in that the image recognition unit introduces an independent parameter update method based on backpropagation of the loss function to optimize and train the visual Transformer encoding network. In this example, the image recognition unit specifically includes the following: constructing a visual Transformer encoding network, and introducing an independent parameter update method based on backpropagation of the loss function during the training process to optimize the visual Transformer encoding network, resulting in a trained visual Transformer encoding network; processing the target region image using the trained visual Transformer encoding network to generate a fixed-dimensional feature vector representation; calculating the similarity of the fixed-dimensional feature vector representation using a metric learning mechanism, and matching it with the historical feature vector library maintained in the storage module to complete the confirmation of the target category attribute and the consistency determination of the target identity features, outputting the identity recognition result corresponding to the target region image.

[0020] Example 4, based on Example 1, uses a consistent target detection CNN model to perform temporal consistency analysis on the preprocessed feature data for continuous frame features, performs target region localization processing, determines the corresponding target region image, and outputs the target detection result corresponding to the target region image. The process specifically includes the following steps: Step S1: Spatiotemporal standard construction: Perform time synchronization, spatial registration and scale normalization on the preprocessed feature data to establish a unified spatiotemporal reference system and form standardized feature data to eliminate the differences in feature distribution under different acquisition nodes, different imaging conditions and different time series. Step S2: Physical Consistency Teacher Model Training: Using the CNN model in the Teacher-CNN-PINN model as the backbone network for image feature extraction, multi-layer convolution, feature mapping, and target region encoding are performed on the standardized feature data to generate high-dimensional feature representations for target localization and category discrimination. A PINN physical consistency constraint module is constructed in the output layer of the CNN model to formalize the evolution of the target region in continuous time and space dimensions into continuously differentiable physical residual operators. A target detection and recognition loss term is introduced, and a joint target loss function is constructed in conjunction with the physical residual operators. This joint target loss function serves as the training objective function for the Teacher-CNN-PINN model and participates in backpropagation optimization. This ensures that both detection accuracy and spatiotemporal consistency constraints are satisfied during the CNN model training phase, resulting in the trained Teacher-CNN-PINN model. The formulas used are as follows: Joint objective loss function formula: ; in, This represents the joint objective loss function of the Teacher-CNN-PINN model. This represents the target detection and recognition loss term of the Teacher-CNN-PINN model, which measures the deviation between the model's detection output for the target region and the corresponding supervised target. Specifically, it is used to constrain the prediction accuracy of the target location coordinates, target category label, and target confidence to ensure the accuracy of the target detection results. This represents the weighting coefficient of the physical consistency constraint, which is used to adjust the degree of influence of the physical consistency constraint in the joint loss function in order to achieve a balance between detection accuracy and spatiotemporal consistency. The physical consistency residual loss term (physical residual operator) is used to characterize the degree of deviation between the evolution of the model output in the continuous time and spatial dimensions and the preset consistency constraints, thereby constraining the rationality of the model output in terms of temporal continuity, spatial smoothness and cross-frame consistency. Step S3: Consistent Representation Feature Generation: Define a continuous consistent representation space. Using the trained Teacher-CNN-PINN model, perform forward inference to uniformly map the model's output results at different time and spatial locations to the continuous consistent representation space, generating continuous consistent representation features for the corresponding target region. The continuous consistent representation features are the soft physical representation results output by the Teacher-CNN-PINN model, used to characterize the evolutionary relationship and consistency features of the target region in continuous time and spatial dimensions. These representation features serve as continuous and differentiable supervisory information, providing consistency constraints for the subsequent knowledge distillation training of the Student model. Continuous consistent representation space: A continuous and differentiable high-dimensional feature representation space used to carry the output results of the Teacher-CNN-PINN model. The representation vectors in this space simultaneously encode the evolutionary relationship and consistency features of the target region in both time and spatial dimensions. Step S4: Lightweight Student Model Construction: Construct a simplified Student-CNN model. The Student-CNN model has fewer network layers, fewer convolutional channels, and a smaller overall parameter scale than the Teacher-CNN-PINN model, thus reducing the model's computational complexity and inference latency. The Student-CNN model does not include the PINN physical consistency constraint module. Its network structure only retains convolutional computation units used to perform target region localization and category discrimination on preprocessed feature data, thereby avoiding the introduction of high-order physical residual calculations and automatic differentiation operations during the inference stage. The target detection capability of the Student-CNN model comes from the subsequent knowledge distillation training process based on KD-PINN. By performing soft-supervised learning on the continuous consistency representation features output by the Teacher-CNN-PINN model, the Student-CNN model implicitly inherits the spatiotemporal consistency features formed by the Teacher model under physical constraints without explicitly introducing a physical constraint module. This maintains the consistency and stability of target detection results in continuous time and space dimensions while maintaining inference efficiency. Step S5: Continuous Distillation Optimization: Based on the continuous and consistent representation space, a two-term coupling loss function is used to distill and optimize the Student-CNN model, thereby transferring the spatiotemporal consistent features learned in the Teacher-CNN-PINN model to the Student-CNN model. The two-term coupling loss function is constructed as follows: When distilling the Student-CNN model, the KD-PINN continuous distillation mechanism is used. Instead of introducing a discrete probability distribution based on softmax, the continuous and consistent representation features output by the Teacher-CNN-PINN model at the corresponding spatiotemporal location are used as the soft supervision target of the Student-CNN model, allowing the distillation process to occur in a continuous and differentiable output space. Based on the consistency constraints between the Teacher-CNN-PINN model and the Student-CNN model in the continuous output space, a continuous distillation loss function is constructed between the Student-CNN model and the Teacher-CNN-PINN model. This continuous distillation loss function is then jointly modeled with the object detection and recognition loss term of the Student-CNN model to obtain the two-term coupling loss function. The formula used is as follows: Formula for loss function in continuous distillation: ; in, This represents the continuous distillation loss function, used to measure the difference between the continuous consistency representation vector output by the Student-CNN model and the continuous consistency representation vector output by the Teacher-CNN-PINN model, so as to realize the transfer of spatiotemporal consistency features in the Teacher model to the Student model; Indicates the distillation sample index. Indicates the number of distilled samples. Indicates the first The spatiotemporal location input variable corresponding to each distillation sample is used to characterize the joint location of the sample in the time and space dimensions. This indicates that the Teacher-CNN-PINN model, when input... The continuous and consistent representation vector output at the corresponding sample is used as the teacher representation in the distillation training. This indicates that the Student-CNN model, when input... The continuous and consistent representation vector output at the corresponding sample is used as the student representation in the distillation training. This represents the squared Euclidean distance between the Teacher representation vector and the Student representation vector, used to characterize the consistency deviation between the two in the continuous and consistent representation space. This indicates that the errors of all distillation samples are summed to obtain an overall measure of distillation error; Formula for the two-term coupling loss function: ; in, This represents the two-term coupling loss function of the Student-CNN model. This represents the weight coefficient of the object detection and recognition loss, which is used to adjust the degree of influence of the object detection and recognition loss term in the overall training process; This represents the target detection and recognition loss term of the Student-CNN model, which measures the deviation between the target location coordinates, target category label, and target confidence output by Student-CNN based on preprocessed feature data and the corresponding supervised target, thereby directly constraining the target detection accuracy of the Student-CNN model; This represents the continuous distillation loss weight coefficient, used to adjust the strength of the continuous distillation constraint during the training process of the Student-CNN model; Step S6: Feature Inference Output: After distillation training is completed, the trained Student-CNN model is obtained. The trained Student-CNN model is used to perform forward inference processing, perform consistent target detection analysis on the target in consecutive frames, automatically locate the target region image corresponding to the target, and output the target detection result corresponding to the target region image.

[0021] Example 5, according to Figure 2 This embodiment is based on Embodiment 4. In this embodiment, the image recognition unit introduces a cross-layer tensor orthogonalization optimization method to perform cross-layer joint constraints and orthogonalization optimization on the parameter update process within the isomorphic coding layer of the visual Transformer coding network, thereby obtaining the trained visual Transformer coding network. The process of processing the target region image through the trained visual Transformer coding network to generate a fixed-dimensional feature vector representation specifically includes the following steps: Step B1: In the visual Transformer encoding network, identify consecutive isomorphic Transformer encoding layers, and determine the linear projection weight matrix for multi-head self-attention calculation in each encoding layer. The set of linear projection weight matrices includes a query matrix, a key matrix, and a value matrix. Perform cross-layer grouping on the isomorphic Transformer encoding layers according to a preset group size to form a layer group index interval. Step B2: During the training phase, the target recognition loss function is introduced to perform backpropagation calculation on the encoding network of the visual Transformer. For each isomorphic encoding layer within the index interval of each layer group, the gradient information corresponding to its linear projection weight matrix is ​​extracted to form a set of gradient matrices divided by layer, which is used for subsequent cross-layer joint optimization processing. Step B3: Perform momentum accumulation processing on the gradient matrix set, and jointly organize the momentum matrices corresponding to each layer according to the layer index order. Introduce the cross-layer momentum tensor construction formula to construct a third-order momentum tensor for cross-layer joint orthogonalization processing. The formula for constructing the cross-layer momentum tensor is: , ; in, This represents the third-order momentum tensor (momentum representation after cross-layer joint organization). Indicates the parameter type identifier. This represents the tensor stacking operator (which stacks multiple matrices of the same shape sequentially into a third-order tensor). , This indicates the starting and ending layer indices of the layer group index range. This represents the momentum matrix corresponding to each isomorphic layer within the layer group interval; express It belongs to a three-dimensional real tensor space: the first dimension has a size of : Number of rows in the matrix; the size of the second dimension is : Number of columns in the matrix; the size of the third dimension is Number of slices per layer; This represents the slice index, used to indicate which momentum matrix slice in the layer group is being retrieved. This represents the tensor slice notation, indicating that the third dimension index is fixed. The two-dimensional slice matrix retrieved at that time; Step B4: Perform a matrix expansion of the third-order momentum tensor along mode-1 to obtain a flattened matrix; apply a gradient orthogonalization operator to the flattened matrix to perform cross-layer joint orthogonalization processing to eliminate redundant overlap of momentum in the main singular direction of different coding layers, obtaining an orthogonalized matrix; perform mode-1 inverse matrix transformation on the orthogonalized matrix to obtain an orthogonalized momentum tensor, using the following formula: mode-1 tensor flattening: ; in, This represents the mode-1 matrix expansion operator, which expands a third-order tensor into a two-dimensional matrix along the first modulus. This represents the flattened matrix after the mode-1 expansion, with the shape as follows: This is used to put cross-layer information into the same matrix space for unified orthogonalization; The mode-1 expansion concatenates the column space directions of each layer along the column dimension, so that the cross-layer related right singular directions can be jointly constrained in the same matrix space. Matrix orthogonalization iteration: ; in, This represents the orthogonalized matrix, with a shape similar to... Consistent, indicating the updated direction matrix after cross-layer joint constraints; The orthogonalization operator is represented by the Newton–Schulz iterative process. The main singular directions are normalized and decorrelated to generate an update direction matrix that satisfies orthogonality constraints, so as to suppress the collapse of the main gradient direction across layers and maintain the tensor diversity of the update direction. Inverse matrix transformation wrapback: ; in, The mode-1 inverse matrix transformation operator is... The inverse mapping is used to fold a flattened matrix back into a third-order tensor in the order it was assembled during expansion. The orthogonalized momentum tensor obtained by the reflection has the following shape: However, its main cross-layer direction has already passed. Subject to joint constraints; generate Figure 2 A schematic diagram comparing the distribution of principal directions before and after orthogonalization of the cross-layer momentum tensor; Figure 2 In the diagram, the horizontal axis represents the singular direction index (Top-k), indicating the position number of the cross-layer momentum tensor in the main singular direction sorting sequence; the vertical axis represents the energy proportion (normalized singular value), reflecting the relative contribution ratio of each main singular direction to the overall momentum update; the blue curve represents the energy proportion of the main direction before orthogonalization, and the orange curve represents the energy proportion of the main direction after orthogonalization; the blue curve represents the energy proportion distribution of the matrix obtained by expanding the cross-layer momentum tensor using mode-1 before orthogonalization in each main singular direction, characterizing the energy concentration of the cross-layer update direction without orthogonalization constraints; the orange curve represents the energy proportion distribution of the matrix obtained by applying the orthogonalization operator to the expanded matrix in each main singular direction, characterizing the energy redistribution state of the cross-layer update direction in the main direction space after orthogonalization. Through comparison... Figure 2 The distribution patterns of the two curves show that before orthogonalization, the update energy was mainly concentrated in a few principal singular directions, while after orthogonalization, the energy distribution of each principal singular direction tended to be balanced, indicating that the correlation between cross-layer momentum update directions was effectively suppressed.

[0022] Step B5: During the training phase, based on the orthogonalized momentum tensor, perform parameter update operations on each isomorphic Transformer encoding layer within the layer group index interval to complete the optimized training of the visual Transformer encoding network and obtain the trained visual Transformer encoding network. Step B6: During the inference phase, the encoding network of the trained visual Transformer is used to perform forward encoding processing on the target region image, outputting a global token feature vector and a local aggregated feature vector respectively; the global feature vector and the local feature vector are subjected to weighted fusion and normalization processing to generate a fixed-dimensional feature vector representation.

[0023] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An intelligent analysis system for police video images based on artificial intelligence, characterized in that: include: Video acquisition module, edge computing module, alarm early warning module, monitoring center and terminal equipment; The video capture module acquires video image data; The edge computing module preprocesses and extracts features from video image data to obtain preprocessed feature data; The edge computing module is equipped with an artificial intelligence analysis module, including a target detection unit, an image recognition unit, and an abnormal behavior analysis unit; The target detection unit uses a consistent target detection CNN model to perform temporal consistency analysis on the features of consecutive frames in the preprocessed feature data, determine the target region image, and output the target detection results; The image recognition unit processes the target region image through the encoding network of the trained visual Transformer to generate a fixed-dimensional feature vector representation; it calculates the similarity of the fixed-dimensional feature vector representation through a metric learning mechanism and matches it with the historical feature vector library maintained in the storage module of the edge computing module to output the identity recognition result. The abnormal behavior analysis unit determines abnormal behavior criteria based on target detection results and identity recognition results, and outputs the abnormal behavior analysis results. The alarm warning module automatically triggers alarms and generates alarm reports based on the results of abnormal behavior analysis; and transmits the alarm reports to the monitoring center and terminal devices to achieve real-time response.

2. The police video image intelligent analysis system based on artificial intelligence according to claim 1, characterized in that: The preprocessed feature data includes spatial features, motion features, color histogram features, edge features, texture features, and illumination features extracted from video image data.

3. The police video image intelligent analysis system based on artificial intelligence according to claim 1, characterized in that: Consistent object detection CNN models include the Teacher-CNN-PINN model and the Student-CNN model.

4. The police video image intelligent analysis system based on artificial intelligence according to claim 1, characterized in that: The image recognition unit generates a fixed-dimensional feature vector representation by: constructing a visual Transformer encoding network, and introducing a cross-layer tensor orthogonalization optimization method during the training process of this encoding network to perform cross-layer joint constraints and orthogonalization optimization on the parameter update process within the isomorphic encoding layer of the visual Transformer encoding network, thereby obtaining the trained visual Transformer encoding network; and processing the target region image through the trained visual Transformer encoding network to generate a fixed-dimensional feature vector representation.

5. The police video image intelligent analysis system based on artificial intelligence according to claim 3, characterized in that: The process of using a consistent object detection CNN model to perform temporal consistency analysis on the features of consecutive frames in the preprocessed feature data to determine the target region image includes the following steps: Step S1: Perform time synchronization, spatial registration and scale normalization on the preprocessed feature data to establish a unified spatiotemporal reference system and form standardized feature data; Step S2: Using the CNN model in the Teacher-CNN-PINN model as the backbone network for image feature extraction, the standardized feature data is processed to generate a high-dimensional feature representation; a PINN physical consistency constraint module is constructed in the output layer of the CNN model, and the evolution law in the continuous time and space dimensions is formalized into a physical residual operator. The target detection and recognition loss term is introduced and combined with the physical residual operator to construct a joint target loss function, which is used as the training objective function of the Teacher-CNN-PINN model and participates in backpropagation optimization to obtain the trained Teacher-CNN-PINN model. Step S3: Define a continuous and consistent representation space, use the trained Teacher-CNN-PINN model to perform forward inference, map to the continuous and consistent representation space, and generate continuous and consistent representation features. Step S4: Perform soft-supervised learning on the continuous consistency representation features so that the Student-CNN model implicitly inherits the spatiotemporal consistency features formed by the Teacher-CNN-PINN model under physical constraints without explicitly introducing a physical constraint module. Step S5: Based on the continuous and consistent representation space, the Student-CNN model is distilled and optimized using a two-term coupling loss function to realize the transfer of the spatiotemporal consistent features learned in the Teacher-CNN-PINN model to the Student-CNN model. Step S6: After the distillation training is completed, the trained Student-CNN model is obtained. The trained Student-CNN model is then used to perform forward inference processing to locate the target region image.

6. The police video image intelligent analysis system based on artificial intelligence according to claim 1, characterized in that: The continuous consistency representation feature is the soft physical representation result output by the Teacher-CNN-PINN model, which is used to characterize the evolutionary relationship and consistency features of the target region in the continuous time and spatial dimensions.

7. The police video image intelligent analysis system based on artificial intelligence according to claim 4, characterized in that: The process by which an image recognition unit generates a fixed-dimensional feature vector representation includes the following steps: Step B1: In the visual Transformer coding network, identify consecutive isomorphic Transformer coding layers and determine the linear projection weight matrix in each coding layer; perform cross-layer grouping on the isomorphic Transformer coding layers according to a preset group size to form layer group index intervals; Step B2: During the training phase, the target recognition loss function is introduced to perform backpropagation calculation on the encoding network of the visual Transformer. For each isomorphic encoding layer within the index interval of each layer group, the gradient information corresponding to its linear projection weight matrix is ​​extracted to form a set of gradient matrices. Step B3: Perform momentum accumulation processing on the gradient matrix set, and jointly organize the momentum matrices corresponding to each layer according to the layer index order. Introduce the cross-layer momentum tensor construction formula to construct a third-order momentum tensor. Step B4: Perform matrix expansion on the third-order momentum tensor along mode-1 to obtain the flattened matrix; apply the gradient orthogonalization operator to the flattened matrix and perform cross-layer joint orthogonalization to obtain the orthogonalized matrix; perform mode-1 inverse matrix transformation on the orthogonalized matrix to obtain the orthogonalized momentum tensor. Step B5: During the training phase, based on the orthogonalized momentum tensor, perform parameter update operations on each isomorphic Transformer encoding layer within the layer group index interval to complete the optimized training of the visual Transformer encoding network and obtain the trained visual Transformer encoding network. Step B6: In the inference phase, the encoding network of the trained visual Transformer is used to perform forward encoding processing on the target region image, outputting global feature vectors and local feature vectors respectively; weighted fusion and normalization processing is performed on the global feature vectors and local feature vectors to generate fixed-dimensional feature vector representations.

8. The police video image intelligent analysis system based on artificial intelligence according to claim 1, characterized in that: The system uses blockchain technology to encrypt and store the police reports generated by the police warning module and their associated target detection results, identity recognition results and abnormal behavior analysis results, and verify their consistency. This ensures the immutability, trustworthiness and compliance of the data, and meets the data privacy and security requirements in the law enforcement process.

Images