A method, system, and device for intelligent identification and early warning of abnormal behavior in a video private network.
By integrating an inertial measurement unit and an image sensor into a video private network, synchronously acquiring and stitching physical pose parameters, performing encrypted transmission with digital signatures, and comparing cross-modal geometric differences at the receiving end, the lack of defense capabilities against forgery and adversarial sample injection at the source level of the video private network is solved, and effective defense against physical field-of-view hijacking attacks is achieved.
Patent Information
- Application Number
- CN202611352950.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-02
- Publication Date
- 2026-11-17
AI Technical Summary
Existing video private networks lack the ability to defend against physical environment replay spoofing and adversarial sample injection at the source level, and commercial cryptographic technologies cannot effectively identify and provide early warnings.
By integrating an inertial measurement unit into the image sensor, physical pose parameters are acquired synchronously, and digital signatures are performed after being stitched together with video frame content. Supplementary enhancement information is encapsulated in the video stream, and encrypted transmission is carried out using a cryptographic processing hardware chip. The receiving end performs signature verification and cross-modal geometric difference comparison to determine the field of view hijacking anomaly.
It effectively defends against physical field-of-view hijacking attacks, ensures the security of video sources, prevents forged images and adversarial sample attacks, and constructs a complete anti-counterfeiting trust loop.
Smart Images

Figure CN123123573A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of image processing and multimodal fusion as well as video network security technology. Specifically, it relates to a method, system, and device for intelligent identification and early warning of abnormal behavior in a video private network. Background Technology
[0002] With the widespread application of mobile vision terminals in special operation scenarios, these devices typically need to transmit real-time footage of on-site operations via a secure and controlled video network. Security attacks targeting these terminals have gradually shifted from eavesdropping at the network transport layer to the physical acquisition layer and the perception algorithm layer. Attackers often employ physical field-of-view hijacking techniques, placing a high-resolution screen in front of the robot's image acquisition sensors to directly play pre-generated fake videos, or injecting minute, imperceptible adversarial noise into the real image, thereby tampering with the intelligent recognition results of subsequent machine vision algorithms at the source of information acquisition.
[0003] The current video private network standard mainly relies on commercial cryptographic technology for digital signatures and link encryption of video signaling and data. This cryptographic system can only defend against packet capture and illegal tampering during network transmission. However, when image forgery or noise injection occurs directly at the physical input stage of the image acquisition component, commercial cryptographic chips will indiscriminately sign and transmit fake video content, causing the entire security trust chain to break at the source. Current anomaly detection technology mainly uses convolutional neural networks to perform pattern matching and recognition of spatial features at the pixel level, lacking cross-dimensional geometric constraint verification of the physical motion state of the device during shooting. When facing advanced deepfake videos with natural motion rendering effects, traditional spatial detection algorithms have a high error detection rate and cannot effectively identify and warn against forgery attacks and adversarial attacks at the source. The existing technology system has a technical deficiency of lacking defense capabilities at the physical input level of the source. Summary of the Invention
[0004] This application addresses the lack of defense capabilities of existing video private networks against physical environment replay spoofing and adversarial sample injection at the source level.
[0005] This application provides a method for intelligent identification and early warning of abnormal behavior in a video private network, comprising: acquiring video streams using an image sensor and simultaneously reading the physical pose parameters output by a rigidly fixed inertial measurement unit during the acquisition phase; calculating the hash digest value of the current video frame content, concatenating it with the extracted physical pose parameters, and digitally signing the concatenated packet using a preset private key through a cryptographic processing hardware chip; encapsulating the signature result and physical pose parameters within a supplementary enhancement information message at the video stream coding level, encrypting the pixel data of the video stream, and transmitting it over the network; stripping the received supplementary enhancement information message and verifying the signature using a public key, and extracting the sequence of consecutive video frames after successful verification; extracting the dense optical flow field matrix of adjacent frames to obtain visual trajectory estimation values; performing filtering and integration on the physical pose parameters to obtain physical reference parameters, and comparing them with the visual trajectory estimation values across modal geometric differences; if the difference parameter is greater than a set boundary threshold, it is determined that there is a physical field-of-view hijacking anomaly.
[0006] Furthermore, the steps of acquiring video streams using an image sensor and simultaneously reading the physical pose parameters output by a rigidly fixed inertial measurement unit during the acquisition phase include: based on the inertial measurement unit rigidly soldered to the center of the back of the printed circuit board of the image sensor, simultaneously acquiring triaxial acceleration data and triaxial angular velocity data that constitute the physical pose parameters; and through a cryptographic processing hardware chip mounted on the internal bus, concurrently capturing video frame pixels of the current exposure period and inertial data buffer blocks with timestamp mapping relationships.
[0007] Further, the step of calculating the hash digest value of the current video frame content and concatenating it with the extracted physical pose parameters includes: calculating the hash digest value of the video frame pixels in the current exposure period based on a standardized general cryptographic algorithm, and concatenating it bit-by-bit with the extracted triaxial acceleration data and triaxial angular velocity data in a memory buffer to form a data block to be signed; using a cryptographic processing hardware chip to call an asymmetric encryption mechanism to generate a signature certificate for the data block to be signed; writing the signature certificate and physical pose parameters into the extended payload of the supplementary enhanced information network abstraction layer unit of the video stream, and using a symmetric encryption mechanism to perform encryption operations on the keyframe image region.
[0008] Furthermore, the steps for extracting the dense optical flow field matrix of adjacent frames to obtain the visual trajectory estimation value include: performing grayscale conversion and pyramid layering on the extracted continuous video frame sequence; extracting pixel-level optical flow vectors based on constant brightness boundaries to establish a dense optical flow field matrix; projecting the dense optical flow field matrix onto three-dimensional space according to the camera intrinsic parameter matrix to establish a visual error cost function; and iteratively updating the three-dimensional rotation matrix component and displacement vector component of the visual trajectory estimation value.
[0009] Furthermore, the steps of performing filtering and integration on the physical pose parameters to obtain physical reference parameters and comparing them with the visually calculated trajectory estimates across modal geometric dissimilarity include: configuring a Kalman filter model and smoothing the received physical pose parameters as observation inputs; performing dual time integration on the smoothed acceleration input to obtain physical reference displacement components, and performing quaternion integration on the smoothed angular velocity input to obtain physical reference rotation components; calculating the first matrix distance between the three-dimensional rotation matrix component and the physical reference rotation component, and the second spatial distance between the displacement vector component and the physical reference displacement component; and comparing the weighted sum of the first matrix distance and the second spatial distance as a dissimilarity parameter with a preset boundary threshold.
[0010] Furthermore, under the condition that the difference parameter is not greater than the set boundary threshold, the following steps are performed: the pixel matrix of the current video frame is subjected to two-dimensional discrete cosine transform using discrete frequency domain analysis logic, and it is mapped from the spatial domain to a two-dimensional frequency distribution matrix containing multiple spatial frequency components; high-frequency coordinate intervals are extracted from the two-dimensional frequency distribution matrix using predefined mask spatial coordinates, and the energy aggregation value within the high-frequency coordinate interval is calculated.
[0011] Furthermore, after calculating the energy aggregation value within the high-frequency coordinate interval, the following steps are performed: calculating the total energy value of all coordinate intervals in the two-dimensional frequency distribution matrix, dividing the energy aggregation value within the high-frequency coordinate interval by the total energy value of the entire frequency band to obtain the high-frequency energy ratio parameter; subtracting the current high-frequency energy ratio parameter from the baseline energy parameter statistically recorded within the preset historical time window to obtain the relative amount of energy surge; when it is determined that the relative amount of energy surge is greater than the critical coefficient for adversarial sample identification, blocking the forwarding process of the current video frame, and generating abnormal interception log data with current clock information and high-frequency coordinate characteristics.
[0012] This application provides an intelligent identification and early warning system for abnormal behavior in a video private network, including a microprocessor and a visual acquisition module, a cryptographic processing hardware chip, and an early warning calculation module connected to the microprocessor. The visual acquisition module includes an image sensor and an inertial measurement unit rigidly connected on the same board, configured to synchronously acquire a continuous video frame sequence and corresponding physical pose parameters. The cryptographic processing hardware chip is configured to calculate the hash digest value of the current video frame content, concatenate it with the extracted physical pose parameters, digitally sign the concatenated packet using a preset private key, encapsulate the signature result and physical pose parameters within a supplementary enhancement information message, and encrypt the pixel data before transmission. The early warning calculation module is configured to strip the received supplementary enhancement information message and verify the signature; after successful verification, extract the continuous video frame sequence; extract the dense optical flow field matrix of adjacent frames to obtain visual trajectory estimation values; perform filtering and integration on the physical pose parameters to obtain physical reference parameters and compare them with the visual trajectory estimation values across modal geometric differences; if the difference parameter is greater than a set boundary threshold, it is determined that there is a physical field-of-view hijacking anomaly.
[0013] Furthermore, the visual acquisition module is equipped with a backplane patch encapsulation layer, and the inertial measurement unit is patch-soldered and mounted on the center area of the back of the printed circuit board of the image sensor; the early warning calculation module is equipped with a nonlinear iterative submodule that iteratively approximates the distribution vector of the spatial rotation matrix and the distribution of the displacement vector to establish the visual trajectory estimation value, as well as a filter bank that performs observation noise suppression operation.
[0014] This application provides a computer device, including a memory and a processor; the memory is configured to store a computer program, and when the processor executes the computer program, it implements any of the above-mentioned intelligent recognition and early warning methods for abnormal behavior of video private networks.
[0015] This application uses underlying hardware to forcibly associate the terminal's actual physical inertial pose information as a reference feature with the hash value of visual data, and encapsulates this association in the transmission message using a cryptographic algorithm, forming a strong, tamper-proof underlying binding. By establishing a cross-modal optical flow consistency solution model at the receiving end, geometric fitting verification between the physical world's motion trajectory and the purely visually inferred trajectory is achieved, preventing physical field-of-view hijacking attacks. Based on frequency domain energy analysis of continuous video frames under legitimate physical motion laws, adversarial sample interference attacks that inject perturbation noise without altering the physical macroscopic motion characteristics are effectively intercepted. The collaborative verification of each stage constructs a complete anti-counterfeiting trust closed loop from physical source input to private network transmission. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the overall structure of the intelligent identification and early warning system for abnormal behavior in a private video network provided in an embodiment of the present invention.
[0017] Figure 2This is a partial cross-sectional view of the vision acquisition module provided in an embodiment of the present invention.
[0018] Figure 3 This is a logical structure block diagram of the early warning calculation module provided in this embodiment of the invention.
[0019] Figure 4 This is a flowchart of the intelligent identification and early warning method for abnormal behavior of video private networks provided in the embodiments of the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. It is to be understood that the specific embodiments described herein are only for explaining related technologies and are not intended to limit the technology. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.
[0021] Example 1: A method for intelligent identification and early warning of abnormal behavior in a video private network like Figure 4 As shown, Figure 4 This is a flowchart of the intelligent identification and early warning method for abnormal behavior in a video private network provided in this embodiment of the invention. The method is applied to the network architecture of a standard video private network and covers the entire process of data acquisition, encryption encapsulation, and cross-modal computation and comparison between the front-end dynamic visual terminal and the back-end private network control center. Specifically, the method includes the following steps: Step S401: Use an image sensor to acquire video stream and read physical pose parameters, and simultaneously read the physical pose parameters output by the rigidly fixed inertial measurement unit during the acquisition phase.
[0022] To ensure strict binding between the optical image and the physical motion state in the time and space coordinate systems, this embodiment imposes strict constraints on the underlying physical grasping process. Based on an inertial measurement unit (IMU) rigidly mounted on the back center of the printed circuit board of the image sensor, three-axis acceleration and three-axis angular velocity data constituting the physical pose parameters are acquired synchronously. The IMU employs a hardware-level six-axis motion processor integrated within the sensor. Through a hard synchronization trigger mechanism using a pin clock signal, the movement attitude of the camera's optical center is synchronously read by the accelerometer and gyroscope within a nanosecond time difference. During this process, the system calls the pre-set intrinsic parameter matrix calibration configuration component and parameter memory module via the internal data bus to synchronously preload the factory-calibrated camera spatial distortion parameters into the memory pool. Through the patch-mount bonding and strong clock binding mode, clutter interference from external mechanical vibrations is effectively eliminated, establishing an observation source with a physical absolute reference anchor point.
[0023] Step S402: Calculate the hash digest value of the current video frame content, concatenate it with the extracted physical pose parameters, and digitally sign the concatenated packet using a preset private key through a cryptographic processing hardware chip.
[0024] After extracting the original visual and inertial signals, the system uses a cryptographic processing hardware chip mounted on the internal bus to simultaneously capture video frame pixels from the current exposure period and inertial data buffer blocks with timestamp mappings. The cryptographic processing hardware chip employs a standardized cryptographic chip. The microprocessor, based on a standardized general cryptographic algorithm, calculates hash digest values for the plaintext pixels of the video frames within the current exposure period, compressing the image spatial dimension into a fixed-length feature fingerprint. In the dedicated chip's memory buffer array, this hash digest value is combined bit-by-bit with the extracted triaxial acceleration and triaxial angular velocity data according to a preset message structure to form a data block to be signed. Using this cryptographic processing hardware chip, an asymmetric encryption mechanism is invoked to read the device's pre-set private key, stored securely and unreadable externally, and a signing operation is performed on the data block to be signed, generating a unique signature certificate. Through hardware-level cryptographic operations, video pixels and the camera's physical motion state are cryptographically fused and bound together.
[0025] Step S403: The signature result and physical pose parameters are encapsulated within the supplementary enhancement information message of the video stream coding layer, and the pixel data of the video stream is encrypted and transmitted over the network.
[0026] To ensure the security payload can successfully pass through packet checks at various routing nodes in an environment compliant with video private network standards, this application employs a deep extension channel for video protocol encapsulation. The controller writes the generated signature certificate and the plaintext data of the physical pose parameters into the extended payload within the Network Abstraction Layer Unit (NALU) of the Supplemental Enhancement Information (SEI) at the video coding level. A symmetric encryption mechanism is used to perform streaming encryption on the pixel data of keyframes and prediction frame image regions of the video stream. After completing source encapsulation and stream encryption, the system calls the network interface hardware and signaling encapsulation underlying driver environment that meet the requirements for stable network stream transmission, sending the integrated hybrid encrypted message across private network nodes to the backend early warning center. This approach of forcibly binding identity verification and physical environment status within the underlying network transmission frame effectively prevents physical separation and fragmentation along the intermediate transmission path.
[0027] Step S404: Strip the received supplementary enhancement information message and verify the signature using the public key, and extract the continuous video frame sequence after the signature is verified.
[0028] On the server side, the private network protocol parsing engine intercepts the incoming data stream and, based on the network abstraction layer unit frame header flag, accurately strips away the supplementary enhancement information messages hidden in the coding layer, extracting the plaintext of the physical pose payload transmitted by the terminal and the digital signature value. The early warning server calls the public key certificate pre-registered and filed with the central node by the corresponding visual terminal to execute the public key verification process. If the verification command fails, it indicates that the source node of the data stream is unauthorized or the message content has been subjected to an illegal man-in-the-middle network layer tampering attack, and the system will immediately cut off the connection channel and discard the abnormal data stream. If the verification command passes, the server will use the negotiated session key to decrypt the video layer, restore and extract the clean continuous video frame sequence, and send it to the buffer queue.
[0029] Step S405: Extract the dense optical flow field matrix of adjacent frames to obtain the visual trajectory estimation value.
[0030] This method reverse-engineers camera motion from purely visual images, employing a computational hardware system that performs heterogeneous matrix calculations and image optical flow processing. It then performs grayscale conversion and Gaussian pyramid layering on the extracted continuous video frame sequences. Based on the assumption of constant brightness and spatial smoothness boundary constraints, it extracts pixel-level optical flow motion vectors by traversing multiple scale levels, establishing a dense optical flow field matrix. This matrix quantitatively characterizes the motion velocity and direction of each corresponding pixel in a two-dimensional visual coordinate system between consecutive frames.
[0031] To construct the mapping relationship from 2D pixel offset to 3D spatial motion, the system projects the dense optical flow field matrix into 3D space based on a preset camera intrinsic parameter matrix, establishing a visual error cost function. This function is used to minimize the projection residual through nonlinear optimization. The system performs the calculation to establish the visual error cost function, using the following formula for iterative minimization:
[0032] Among them, in the formula J This represents the total visual error cost function value, which indicates the degree of fit between the calculated result and the actual pixel deflection. This indicates the first video frame obtained through feature matching. i The two-dimensional pixel coordinates of each tracking pixel on the normalized image plane; This indicates the corresponding number in the current video frame. i The coordinates of each matching pixel; K This represents the camera intrinsic parameter matrix read from the parameter memory, which determines the focal length and principal point offset properties of the optical lens. It is the inverse of the intrinsic parameter matrix; This represents the estimated value of the three-dimensional rotation matrix components in the visually inferred trajectory estimate of the target obtained through iterative solution; This represents the displacement vector estimate in the visual trajectory estimate of the target being solved; Z i This represents the depth information of the corresponding spatial feature points based on depth estimation or binocular baseline extrapolation. The system employs a nonlinear least squares method to iteratively update and solve this function to obtain visual trajectory estimates with minimal local errors.
[0033] Step S406: Filter and integrate the physical pose parameters to obtain physical reference parameters, and compare them with the visually calculated trajectory estimate across modal geometric differences; if the difference parameter is greater than the set boundary threshold, it is determined that there is a physical field of view hijacking anomaly.
[0034] Since the physical pose parameters obtained through decryption inevitably incorporate sensor thermal noise and high-frequency jitter components during acquisition, smoothing processing must be performed first. The system is configured with a Kalman filter model to smooth and denoise the received triaxial acceleration and triaxial angular velocity physical pose parameters as observation inputs. For the smoothed angular velocity input, the system performs quaternion-based ordinary differential equation integration to obtain the physical reference rotation component. R t Perform dual-time continuous integration calculations on the smoothed acceleration input to obtain the physical reference displacement components. T t .
[0035] The system calculates the first matrix distance between the three-dimensional rotation matrix components and the physical reference rotation components, and the second spatial distance between the displacement vector components and the physical reference displacement components. These two distances are then weighted and summed according to preset weights. The system then calculates the cross-modal difference index parameters. The calculation formula is as follows:
[0036] in, A parameter representing the cross-modal dissimilarity index indicating the fusion of multi-dimensional spatial motion characteristics; The Frobenius norm operator represents the matrix and is used to measure the Euclidean space scale of different rotation matrices; and These are the estimated values of the three-dimensional rotation matrix components and displacement vector obtained from the aforementioned visual optical flow calculation; R t T The transpose of the actual physical reference rotation component matrix; I It is the identity matrix; T t For actual physical reference displacement components; and These are weighting coefficients preset based on the stiffness and vibration characteristics of the specific robot's suspension system. The preferred value range is 0.5 to 0.7. The preferred value range is 0.3 to 0.5, which balances the influence of rotation error and translation error on anomaly determination.
[0037] In real-time stream comparison, the system determines the calculated cross-modal difference index parameters. If the value exceeds a set threshold, it indicates a significant geometric deviation between the spatial relative motion trend shown in the image sequence and the actual physical vibration and deflection trend recorded by the gyroscope mounted on the back of the camera. Based on this, the system determines that it is under attack at the source level outside the network layer, i.e., an abnormal physical field-of-view hijacking, where a high-definition flat panel display is placed in front of the robot's camera to record and replay forged images. Through multi-dimensional cross-verification combining hardware base source binding and spatial geometric alignment, the system intercepts replay video footage that is legitimate at the commercial cryptographic protocol layer but has a false physical content.
[0038] Step S407: Under the condition that the difference parameter is not greater than the set boundary threshold, the discrete frequency domain analysis logic is used to verify and intercept the anomaly.
[0039] To counter advanced network attack patterns such as adversarial noise injection, attackers do not forge or tamper with macroscopic scene motion trajectories. Instead, they inject high-frequency perturbation noise into the pixels of real video frames, misleading the backend machine vision recognition algorithm. Under the condition that the difference parameter does not exceed a set boundary threshold, the system invokes discrete frequency domain analysis logic to perform a two-dimensional discrete cosine transform on the current clean front-end video frame pixel matrix channels, mapping it from the original two-dimensional spatial domain to a two-dimensional frequency distribution matrix containing multiple spatial frequency components.
[0040] The system performs calculations to establish a two-dimensional discrete cosine transform spectrum derivation model, and obtains the amplitude value at any frequency coordinate point within the matrix using the following formula:
[0041] in, D(u,v) This indicates the row frequency in the two-dimensional frequency distribution matrix obtained after the transformation. u Sum column frequency v The discrete frequency domain band component energy amplitude parameter at the location; V t (x,y) This indicates the spatial coordinates position in the original current plaintext video frame. (x,y) The grayscale value parameter of the pixel brightness channel at that location; N This represents the cardinality of the width and height dimensions of the macroblock window used for image slice analysis, with the preferred size being... or Pixel boundary values; c(u) and c(v) All are orthogonal transformation normalization compensation coefficients.
[0042] The system uses predefined mask space coordinates to extract the high-frequency coordinate interval located in the lower right corner of the two-dimensional frequency distribution matrix. It then calculates the aggregated energy value within the high-frequency coordinate interval by squaring and summing the amplitude matrix. Finally, the system calculates the total full-band energy value for all coordinate intervals in the two-dimensional frequency distribution matrix and divides the aggregated energy value within the high-frequency coordinate interval by the total full-band energy value to obtain the high-frequency energy proportion parameter.
[0043] The system maintains a sliding historical queue, subtracting the current high-frequency energy percentage parameter from the baseline energy parameter within a preset historical time window to obtain the relative amount of energy surge. When the relative amount of energy surge is determined to be greater than the adversarial example identification threshold, an abnormal surge in the energy percentage of the high-frequency region of the frequency band is detected, triggering the security baseline. The system blocks the input of the current video frame sequence containing malicious features and its subsequent flow process to prevent contamination of the downstream target detection neural network engine, and generates abnormal interception log data with current clock information and high-frequency coordinate features for archiving and auditing. Through this frequency domain collaborative filtering mechanism, interference that does not change the physical motion laws but disrupts the continuous microscopic features of pixels is intercepted.
[0044] Example 2: An Intelligent Recognition and Early Warning System for Abnormal Behavior in a Private Video Network This embodiment provides an intelligent identification and early warning system for abnormal behavior of a video private network for implementing the method described in Embodiment 1.
[0045] Figure 1 This is a schematic diagram of the overall structure of the intelligent identification and early warning system for abnormal behavior in a video private network provided in this embodiment of the invention. The system mainly comprises the visual acquisition device located at the field operation end and the aggregation computing node located at the back-end control center. The core control hub of the system includes a microprocessor 101, which establishes stable communication connections with the visual acquisition module 102, the cryptographic processing hardware chip 103, and the remotely deployed early warning calculation module 104 through an internal data interaction bus and power supply circuit. The early warning calculation module 104 is mainly deployed inside the early warning server node supported by computing power.
[0046] Figure 2This is a partial cross-sectional view of the visual acquisition module 102 provided in an embodiment of the present invention. The visual acquisition module 102 includes an image sensor 201 for optical imaging and a printed circuit board 202. In this module, a backplane patch encapsulation layer 204 is configured, and an inertial measurement unit 203 is patch-soldered and tightly mounted to the central region of the back of the printed circuit board 202 of the image sensor 201. This microscopic, rigid connection assembly process ensures that, during movement or vibration, the angular velocity and acceleration states recorded by the inertial measurement unit 203 can represent the physical spatial pose changes of the lens of the image sensor 201 without delay or flexible deformation. The visual acquisition module 102 is configured in this structure to synchronously acquire continuous video frame sequences and corresponding physical pose parameters.
[0047] The cryptographic processing hardware chip 103 adopts a cryptographic hardware chip that conforms to the national cryptographic standards. This cryptographic processing hardware chip 103 is configured and executes core operational logic instructions, reads the plaintext of the video frame transmitted from the microprocessor 101, calculates the hash digest value of the current video frame content, directly concatenates it with the extracted physical pose parameters in a secure memory area, and performs a high-security digital signature on the concatenated packet using a pre-set private key protected by the hardware physical layer; encapsulates the signature result and the plaintext of the physical pose parameters within a supplementary enhanced information message at the encoding level, and encrypts the pixel backbone data before transmitting it over the network.
[0048] Figure 3 This is a logical structure block diagram of the early warning resolution module 104 provided in this embodiment of the invention. The early warning resolution module 104 is built into a heterogeneous computing hardware platform in the central computer room. Its overall configuration is to complete the protocol reverse decoding and stripping of received supplementary and enhanced information messages and perform high-strength public key verification; after the verification command passes, it extracts continuous video sequences without loss. Logically, the early warning resolution module 104 includes an optical flow calculation and analysis module 301, a matrix geometry resolution module 302, and a discrete frequency domain analysis module 303. The optical flow calculation and analysis module 301 extracts the dense optical flow field matrix of adjacent frames through a parallel pipeline algorithm and obtains the visual trajectory estimation value through a built-in nonlinear iterative submodule. The matrix geometry solution module 302 is equipped with a Kalman and integral filter bank to perform observation noise suppression operations. It performs filtering and integration on the physical pose parameters to obtain physical reference parameters and compares them with the visual trajectory estimation value output by the optical flow calculation and analysis module 301 for cross-modal geometric difference. If the difference parameter output by the solution is greater than the set boundary threshold, an alarm signal indicating the presence of a physical field-of-view hijacking anomaly is output to the central control console. For video stream segments that do not show physical hijacking anomalies, they are distributed to the discrete frequency domain analysis module 303 to perform two-dimensional high-frequency energy surge characteristic verification and analysis operations.
[0049] The system possesses the ability to resist external physical screen copying and forgery, as well as the ability to use invisible anti-sample noise intrusion, thus ensuring the source security of the transmission link of the private network intelligent video terminal.
[0050] Example 3: A computer device This invention also provides a computer device, including a memory and a processor, which are hardwired together via a bus or other communication protocol. The memory, as a computer-readable storage medium, is configured to store computer program instructions. The processor executes the various steps of the intelligent identification and early warning method for abnormal behavior in a video private network described in Embodiment 1 by running or executing software program instructions and data modules stored in the memory. The processor drives the device to perform actions such as reading timestamp-aligned video frames and sensor parameters, generating digital signature spliced messages, constructing optical flow iterative equations, performing two-dimensional frequency distribution feature mapping comparison, and intercepting and outputting alarms. In actual industrial deployments, this computer device is deployed as an edge computing node.
[0051] The above description is merely a specific embodiment of this application, and the scope of protection of this application is not limited thereto. Any variations or substitutions that can be conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. The scope of protection of this application shall be determined by the scope of the claims.
Claims
1. A method for intelligent identification and early warning of abnormal behavior in a video private network, characterized in that, include: The video stream is acquired using an image sensor, and the physical pose parameters output by the rigidly fixed inertial measurement unit are read synchronously during the acquisition phase. Calculate the hash digest value of the current video frame content, concatenate it with the extracted physical pose parameters, and digitally sign the concatenated packet using a preset private key through a cryptographic processing hardware chip; The signature result and physical pose parameters are encapsulated within the supplemental enhancement message at the video stream coding level, and the pixel data of the video stream is encrypted and transmitted over the network. The received supplemental enhancement information message is stripped and verified using the public key. After successful verification, the sequence of consecutive video frames is extracted. Extract the dense optical flow field matrix of adjacent frames to obtain the visual trajectory estimation value; Filtering and integration are performed on the physical pose parameters to obtain physical reference parameters, and these are compared with the visually inferred trajectory estimates across modal geometric differences. If the difference parameter is greater than the set boundary threshold, it is determined that there is a physical field of view hijacking anomaly.
2. The method as described in claim 1, characterized in that, The steps for acquiring video streams using an image sensor and simultaneously reading the physical pose parameters output by a rigidly fixed inertial measurement unit during the acquisition phase include: Based on the inertial measurement unit rigidly attached to the center of the back of the printed circuit board of the image sensor, the three-axis acceleration data and three-axis angular velocity data that constitute the physical pose parameters are acquired simultaneously. The cryptographic processing hardware chip mounted on the internal bus captures video frame pixels of the current exposure period and inertial data buffer blocks with timestamp mapping relationships in parallel.
3. The method as described in claim 2, characterized in that, The steps of calculating the hash digest value of the current video frame content and concatenating it with the extracted physical pose parameters include: The hash digest value of the video frame pixels in the current exposure period is calculated based on a standardized general cryptographic algorithm, and then concatenated bit by bit with the extracted triaxial acceleration data and triaxial angular velocity data in the memory buffer to form a data block to be signed. A signature certificate is generated by using a cryptographic processing hardware chip to retrieve an asymmetric encryption mechanism from the data block to be signed. The signature certificate and physical pose parameters are written into the supplemental and enhanced information network abstraction layer unit of the video stream, and the keyframe image region is encrypted using a symmetric encryption mechanism.
4. The method as described in claim 1, characterized in that, The steps for extracting the dense optical flow field matrix of adjacent frames to obtain the visual trajectory estimate include: The extracted video continuous frame sequence is subjected to grayscale conversion and pyramid layering, and a dense optical flow field matrix is established by extracting pixel-level optical flow vectors based on constant brightness boundaries. Based on the camera intrinsic parameter matrix, the dense optical flow field matrix is projected onto the three-dimensional space to establish a visual error cost function, and the three-dimensional rotation matrix components and displacement vector components of the visual trajectory estimation are iteratively updated.
5. The method as described in claim 4, characterized in that, The steps of performing filtering and integration on the physical pose parameters to obtain physical reference parameters and comparing them with the visually inferred trajectory estimates across modal geometric differences include: Configure a Kalman filter model to smooth the received physical pose parameters as observation inputs; The smoothed acceleration input is subjected to dual time integration to obtain the physical reference displacement component, and the smoothed angular velocity input is subjected to quaternion integration to obtain the physical reference rotation component. Calculate the first matrix distance between the three-dimensional rotation matrix components and the physical reference rotation components, and the second spatial distance between the displacement vector components and the physical reference displacement components; The weighted sum of the first matrix distance and the second spatial distance is used as the difference parameter and compared with a preset boundary threshold.
6. The method as described in claim 1, characterized in that, If the difference parameter is not greater than the set boundary threshold, the following steps are performed: The pixel matrix of the current video frame is subjected to a two-dimensional discrete cosine transform using discrete frequency domain analysis logic, which maps it from the spatial domain to a two-dimensional frequency distribution matrix containing multiple spatial frequency components. High-frequency coordinate intervals are extracted from the two-dimensional frequency distribution matrix using predefined mask space coordinates, and the energy aggregation value within the high-frequency coordinate intervals is calculated.
7. The method as described in claim 6, characterized in that, After calculating the energy aggregation value within the high-frequency coordinate range, the following is also performed: Calculate the total energy value of the entire frequency band in all coordinate intervals of the two-dimensional frequency distribution matrix, and divide the aggregated energy value in the high-frequency coordinate interval by the total energy value of the entire frequency band to obtain the high-frequency energy ratio parameter; The relative amount of energy surge is obtained by subtracting the current high-frequency energy percentage parameter from the baseline parameter of the basic energy statistically collected within the preset historical time window. When the relative amount of energy surge is determined to be greater than the critical coefficient for adversarial example identification, the forwarding process of the current video frame is blocked, and abnormal interception log data with current clock information and high-frequency coordinate features is generated.
8. A smart identification and early warning system for abnormal behavior in a video private network, characterized in that, It includes a microprocessor and a vision acquisition module, a cryptographic processing hardware chip, and an early warning calculation module connected to the microprocessor; The visual acquisition module includes an image sensor and an inertial measurement unit rigidly connected to the same board, configured to synchronously acquire a continuous sequence of video frames and the corresponding physical pose parameters; The cryptographic processing hardware chip is configured to calculate the hash digest value of the current video frame content, concatenate it with the extracted physical pose parameters, digitally sign the concatenated packet with a preset private key, encapsulate the signature result and physical pose parameters in a supplementary and enhanced information message, and encrypt the pixel data before transmission. The early warning calculation module is configured to strip the received supplementary enhancement information message and perform signature verification; after the signature verification is passed, extract the continuous video frame sequence; extract the dense optical flow field matrix of adjacent frames to obtain the visual trajectory estimation value; The physical pose parameters are filtered and integrated to obtain physical reference parameters, which are then compared with the visually calculated trajectory estimates across modal geometric differences. If the difference parameter is greater than the set boundary threshold, a physical field of view hijacking anomaly is determined to exist.
9. The system as described in claim 8, characterized in that, The visual acquisition module is equipped with a backplane patch encapsulation layer, and the inertial measurement unit is patch-soldered and mounted on the center area of the back of the printed circuit board of the image sensor; the early warning calculation module is equipped with a nonlinear iterative submodule that iteratively approximates the distribution vector of the spatial rotation matrix and the distribution of the displacement vector to establish the visual trajectory estimation value, as well as a filter bank that performs observation noise suppression operation.
10. A computer device, comprising a memory and a processor, characterized in that, The memory is configured to store a computer program, which, when executed by the processor, implements the method as described in any one of claims 1 to 7.
