Traffic video intelligent desensitization processing method and system

Abstract

The application discloses a kind of traffic video intelligent desensitization processing method and system, it is related to computer vision technical field, including, based on encrypted video data packet, the multistage desensitization channel including pixel layer processing, style layer processing and semantic layer processing is established;Through multi-modal fusion recognition model, feature extraction and semantic analysis are carried out to encrypted video data packet, the identification confidence of sensitive element and context information are acquired, and hierarchical determination is executed, and channel selection instruction is generated;The structural characteristics of original video frame and desensitized video frame are compared, the structural difference degree index is calculated, and structure consistency verification information is generated;Desensitized video frame, structure consistency verification information and sensitive area separation report are uniformly encapsulated, and desensitized data packet is formed.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an intelligent desensitization processing method and system for traffic videos. Background Technology

[0002] With the rapid development of applications such as intelligent transportation, urban safety supervision, and intelligent driving, video acquisition equipment has been widely deployed in the field of road traffic for real-time monitoring, event recognition, and traffic flow analysis. Traditional traffic video processing relies on centralized servers for the acquisition, storage, and analysis of video data. In recent years, with the development of deep learning and multimodal recognition technologies, traffic videos can not only identify targets such as vehicles, pedestrians, and traffic lights, but also analyze traffic congestion trends and abnormal behaviors through spatiotemporal modeling.

[0003] Existing technologies still have limitations in the verifiability of collaborative processing and desensitization of multimodal sensitive information. Current methods mostly focus on pixel-level masking or encrypted transmission, lacking the ability to dynamically resolve the semantic context of sensitive elements (such as behavioral correlation and spatiotemporal attributes), resulting in insufficient adaptability of desensitization strategies. Existing solutions typically do not provide a consistency verification mechanism between the desensitized data and the original data, making it difficult for data users to assess the impact of the desensitization process on the structural integrity of the video, thus limiting the usability of desensitized data in high-precision analysis tasks (such as traffic modeling and accident tracing). Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent desensitization processing method for traffic videos to solve the problems of insufficient adaptability to the hierarchical classification of multimodal sensitive information and lack of process verifiability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent desensitization processing method for traffic videos, comprising: acquiring the original video stream, encrypting and transmitting it to a trusted data space through a privacy protection gateway, and simultaneously attaching spatiotemporal metadata to generate an encrypted video data packet; establishing a multi-level desensitization channel based on the encrypted video data packet, including pixel-level processing, style-level processing, and semantic-level processing; performing feature extraction and semantic analysis on the encrypted video data packet through a multimodal fusion recognition model to obtain the recognition confidence and context information of sensitive elements, and performing hierarchical judgment to generate a channel selection instruction; matching the corresponding processing path in the multi-level desensitization channel according to the channel selection instruction, performing corresponding desensitization on different sensitive areas, and outputting desensitized video frames and sensitive area separation reports; comparing the structural features of the original video frame and the desensitized video frame, calculating the structural difference index, and generating structural consistency verification information; and uniformly encapsulating the desensitized video frame, structural consistency verification information, and sensitive area separation report to form a desensitized data packet.

[0008] In a preferred embodiment of the intelligent desensitization processing method for traffic videos described in this invention, the specific steps for generating encrypted video data are as follows:

[0009] The original video stream is acquired and embedded with six-dimensional spatiotemporal metadata generated by satellite positioning to form a spatiotemporal tensor data packet. The pixel spatial topology of the original video frame is used to construct a dynamic Riemannian manifold and perform nonlinear distortion encryption to obtain the encrypted spatiotemporal tensor data packet.

[0010] The encrypted spatiotemporal tensor data packet is divided into frame units, each frame unit is digitally signed and a Merkle tree is constructed, the root value of the Merkle tree is generated, and the geometric topological consistency of each frame unit with the dynamic Riemannian manifold is verified through a zero-knowledge proof protocol to generate a verified encrypted data stream.

[0011] The verified encrypted data stream is reassembled into a first chain and a second chain according to the parity of frame units, and a manifold connectivity topology certificate is attached to obtain encrypted video data packets.

[0012] As a preferred embodiment of the intelligent desensitization processing method for traffic videos described in this invention, the step of establishing a multi-level desensitization channel based on encrypted video data packets, including pixel-level processing, style-level processing, and semantic-level processing, comprises the following specific steps:

[0013] The frame sequence of encrypted video data packets is parsed to extract spatial feature data, pixel distribution features, time series features, semantic tag features, and object detection results;

[0014] Based on spatial feature data and pixel distribution features, a pixel layer processing channel is constructed through an adaptive blur algorithm and an edge-preserving filtering mechanism. Based on spatial feature data and time series features, a style layer processing channel is constructed through an image style transfer algorithm and a brightness equalization mechanism.

[0015] Based on semantic label features and object detection results, a semantic layer processing channel is constructed through semantic segmentation algorithm and feature alignment mechanism. The pixel layer processing channel, style layer processing channel and semantic layer processing channel are connected in series to form a multi-level desensitization processing channel.

[0016] In a preferred embodiment of the intelligent desensitization processing method for traffic videos described in this invention, the specific steps for generating the channel selection instruction are as follows:

[0017] The spatial distribution features and texture information of encrypted video data packets are extracted by the visual feature extraction branch of the multimodal fusion recognition model to generate a spatial feature set, and the temporal feature branch is extracted by the motion trajectory and inter-frame change information of encrypted video data packets to generate a temporal feature set.

[0018] The semantic understanding branch of the multimodal fusion recognition model is used to extract contextual features and object correlation of encrypted video data packets to generate a semantic feature set;

[0019] The feature fusion layer of the multimodal fusion recognition model fuses spatial feature sets, temporal feature sets, and semantic feature sets to generate a multimodal fusion feature tensor. Then, through the confidence calculation mechanism and context weight function, recognition confidence and context information are generated.

[0020] Based on the recognition confidence level and contextual information, the desensitization level of sensitive targets is determined by the hierarchical judgment logic of the multimodal fusion recognition model, and the channel selection instruction is output.

[0021] In a preferred embodiment of the intelligent desensitization processing method for traffic videos described in this invention, the specific steps for outputting the desensitized video frame and the sensitive region separation report are as follows:

[0022] Select the corresponding desensitization processing channel according to the channel selection instruction, and assemble the corresponding desensitization processing channel and the multimodal fusion feature tensor to generate the data stream to be desensitized;

[0023] Extract the coordinate information of the sensitive area from the data stream to be de-identified, perform the de-identification operation, generate the de-identified feature frame, verify the integrity of the de-identified area boundary through the topology consistency comparison algorithm, and output the corrected de-identified feature frame;

[0024] The frame sequence reconstruction process restores the corrected desensitized feature frame order, generates desensitized video frames, and aggregates the type, location, and processing intensity of sensitive areas to generate a sensitive area separation report.

[0025] In a preferred embodiment of the intelligent desensitization processing method for traffic videos described in this invention, the specific steps for generating structural consistency verification information are as follows:

[0026] Based on the image feature data of the original video frames and the desensitized video frames, edge distribution features, texture gradient features and motion trajectory features are obtained through feature extraction algorithms to obtain a multi-dimensional feature set.

[0027] The structural difference index is calculated using the structural similarity calculation formula and the optical flow consistency function. The structural difference index is then mapped to the structural preservation level. The structural difference index and the structural preservation level are jointly encapsulated to generate structural consistency verification information.

[0028] In a preferred embodiment of the intelligent de-identification processing method for traffic videos according to the present invention, the specific steps for forming the de-identified data packet are as follows:

[0029] Based on desensitized video frames, structural consistency verification information, and sensitive area separation reports, a corresponding relationship is established through a spatiotemporal label matching mechanism to generate an encapsulation index table, which is then mapped and integrated to generate a data fusion structure.

[0030] Based on the data fusion structure, a unique verification code is generated through hash encryption and timestamp identification mechanism. The unique verification code and the encapsulation index table are serialized and combined to generate a de-identified data packet.

[0031] Secondly, this invention provides an intelligent desensitization processing system for traffic videos, comprising: an acquisition module for acquiring raw video streams, encrypting and transmitting them to a trusted data space via a privacy protection gateway, and simultaneously attaching spatiotemporal metadata to generate encrypted video data packets; a construction module for establishing multi-level desensitization channels based on the encrypted video data packets, including pixel-level processing, style-level processing, and semantic-level processing; an instruction module for performing feature extraction and semantic analysis on the encrypted video data packets using a multimodal fusion recognition model, obtaining the recognition confidence and context information of sensitive elements, performing hierarchical judgment, and generating channel selection instructions; a desensitization module for matching corresponding processing paths in the multi-level desensitization channels according to the channel selection instructions, performing corresponding desensitization on different sensitive areas, and outputting desensitized video frames and sensitive area separation reports; a verification module for comparing the structural features of the original video frames and the desensitized video frames, calculating structural difference indices, and generating structural consistency verification information; and an encapsulation module for uniformly encapsulating the desensitized video frames, structural consistency verification information, and sensitive area separation reports to form desensitized data packets.

[0032] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent desensitization processing method for traffic videos as described in the first aspect of the present invention.

[0033] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent desensitization processing method for traffic videos as described in the first aspect of the present invention.

[0034] The beneficial effects of this invention are as follows: by constructing a multi-level desensitization channel, it achieves refined and adaptive hierarchical processing of multimodal sensitive information, solving the problem that the existing desensitization strategies are singular and cannot balance privacy protection and data availability; under the premise of thoroughly protecting privacy, it significantly enhances the value and application potential of desensitized data in high-precision traffic analysis. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of an intelligent desensitization processing method for traffic videos.

[0037] Figure 2 This is a schematic diagram of an intelligent desensitization processing system for traffic videos.

[0038] Figure 3 A flowchart for generating encrypted video data packets.

[0039] Figure 4 A flowchart for constructing a multi-level desensitization processing channel. Detailed Implementation

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0043] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an intelligent desensitization processing method for traffic videos, comprising the following steps:

[0044] S1. Acquire the raw video stream, encrypt and transmit it to a trusted data space through a privacy protection gateway, and simultaneously attach spatiotemporal metadata to generate encrypted video data packets.

[0045] The original video stream is acquired and embedded with six-dimensional spatiotemporal metadata generated by satellite positioning to form a spatiotemporal tensor data packet. The pixel space topology of the original video frame is used to construct a dynamic Riemannian manifold and perform nonlinear distortion encryption to obtain the encrypted spatiotemporal tensor data packet.

[0046] Specifically, the raw video stream is acquired, and the RGB pixel matrix and frame rate parameters of each frame are parsed using the H.264 encoding standard. Simultaneously, six-dimensional spatiotemporal metadata generated by satellite positioning is received, including three-dimensional geographic coordinates (latitude, longitude, altitude) and three-dimensional time parameters (UTC time, leap second correction value, time zone offset). This six-dimensional spatiotemporal metadata is embedded into the raw video stream through the H.265 Extended Supplemental Information (SEI) layer, forming a spatiotemporal tensor data packet. For each frame of the raw video stream in the spatiotemporal tensor data packet, a pixel spatial adjacency graph is calculated: defining the image... Elementary points are topological nodes, and the weights of the edges connecting adjacent pixels are the Euclidean distances of the RGB three channels, constructing a pixel space topological adjacency matrix. A dynamic Riemannian manifold is generated based on the adjacency matrix, and the metric tensor is set as the Jacobian matrix of the pixel color difference gradient function. The curvature change of the pixel space is described by solving the geodesic equation, and nonlinear distortion encryption is achieved by applying chaotic mapping. The pixel displacement is driven by the Lorentz attractor differential equation, and the pixel position offset is obtained iteratively at each discrete time step, so that the pixel migrates along the geodesic path of the manifold, and the encrypted spatiotemporal tensor data packet is output.

[0047] The encrypted spatiotemporal tensor data packet is divided into frame units. Each frame unit is digitally signed and a Merkle tree is constructed. The root value of the Merkle tree is generated. The geometric topological consistency of each frame unit with the dynamic Riemannian manifold is verified through a zero-knowledge proof protocol, and a verified encrypted data stream is generated.

[0048] Specifically, the encrypted spatiotemporal tensor data packet is divided into independent frame units according to the video frame time boundaries. Each frame unit contains encrypted image data with an example value of 1920×1080 pixels and corresponding six-dimensional spatiotemporal metadata. An EdDSA (Edwards Curve Digital Signature Algorithm) digital signature operation is performed on each frame unit: the SHA3-512 hash value of the frame unit data is calculated using the private key to generate a 64-byte signature body. A Merkle binary tree structure is constructed based on all frame unit signatures: the signature of each frame unit is used as a leaf node, and the signatures of adjacent nodes are recursively calculated. The SHA3-256 hash value is used to generate the parent node, resulting in a fixed-length 32-byte Merkle root value, which serves as the cryptographic anchor for the integrity of the entire data packet. The Groth16 zero-knowledge proof protocol is used to verify the geometric topological consistency between the frame unit and the dynamic Riemannian manifold: the pixel spatial topological adjacency matrix within the frame unit is extracted, the Frobenius norm difference value between the frame unit and the initial dynamic Riemannian manifold metric tensor is calculated (e.g., difference value ≤ 0.01), and the proof parameters and verification key are generated. After all frame units pass the verification, the verified encrypted data stream is output.

[0049] It should be noted that the expression for calculating the Frobenius norm difference value is as follows:

[0050]

[0051] in, This represents the Frobenius norm difference value. The Jacobian matrix represents the differential homeomorphism mapping. Represents the identity matrix. This represents the Frobenius norm.

[0052] The verified encrypted data stream is reassembled into a first chain and a second chain according to the parity of frame units, and a manifold connectivity topology certificate is attached to obtain encrypted video data packets.

[0053] Specifically, the verified encrypted data stream is separated according to the parity of the frame sequence: all odd-numbered frame units are extracted to form the first chain, and all even-numbered frame units are formed to form the second chain; each frame unit maintains its original structure (including the encrypted pixel matrix, six-dimensional spatiotemporal metadata, and EdDSA signature), and the inter-chain association constraints are generated by merging the odd and even frame position index matrices through Hadamard product; the manifold connectivity index is calculated based on the initial metric tensor of the dynamic Riemannian manifold: the one-dimensional Betti number of the pixel space topology is solved, the number of connected branches of the road structure is quantified (e.g., the original manifold Betti number = 3) and compared with the encrypted manifold Betti number. When the absolute difference of the example value is ≤ 0.01, a manifold connectivity topology certificate is generated. The certificate contains the initial Betti number, the encrypted Betti number, and the UTC timestamp; the first chain, the second chain, and the manifold connectivity topology certificate are encapsulated, with the first chain storing the odd-numbered frame unit sequence and the second chain storing the even-numbered frame unit sequence. The certificate is appended to the end of the data packet to obtain the encrypted video data packet.

[0054] S2. Based on encrypted video data packets, establish a multi-level desensitization channel that includes pixel layer processing, style layer processing, and semantic layer processing.

[0055] The frame sequence of encrypted video data packets is parsed to extract spatial feature data, pixel distribution features, time series features, semantic tag features, and object detection results.

[0056] Specifically, from the encrypted video data packets recombined according to parity and with attached manifold connectivity topology certificates, the data is read and decrypted frame by frame to recover a processable video frame sequence. Feature extraction algorithms (such as convolution operations or optical flow methods) are used to obtain spatial feature data describing the spatial structure relationship of the image, pixel distribution features reflecting the statistical characteristics of pixel values, motion trajectories and inter-frame change information that characterize the continuous changes between adjacent frames as time series features from each decrypted frame.

[0057] Semantic label features are extracted from the video frame sequence. The scene category, object category and location in each frame are identified by semantic segmentation and classification algorithms to form semantic label features. Object detection is performed on the video frame sequence. The object category, bounding box location and confidence value in each frame are identified by object detection algorithms and associated with the semantic label features to form object detection results.

[0058] Based on spatial feature data and pixel distribution features, a pixel layer processing channel is constructed using an adaptive blur algorithm and an edge-preserving filtering mechanism. Based on spatial feature data and time series features, a style layer processing channel is constructed using an image style transfer algorithm and a brightness equalization mechanism.

[0059] Specifically, spatial feature data (such as tensors describing the contours and relative positions of objects) and pixel distribution features (such as histogram statistics of pixel values) are input into an adaptive blurring algorithm (such as a filter function that dynamically adjusts the Gaussian kernel size based on the gradient magnitude in the pixel distribution features) for processing. The blur intensity is adaptively selected based on the local variance example value of 0.1 to 0.5 in the pixel distribution features. At the same time, an edge-preserving filtering mechanism (such as bilateral filtering based on local gradient direction weights) is combined to specifically preserve high-frequency edge information in the spatial feature data, generating pixel layer processing channels.

[0060] Spatial feature data (such as scene depth information or segmentation mask) and temporal series features (such as optical flow vector field between consecutive frames) are jointly input into an image style transfer algorithm (such as a neural network style transfer method based on Gram matrix statistical feature matching). The temporal series features are used to ensure the consistency of style transformation between adjacent frames to avoid flickering. At the same time, the global illumination distribution of the transferred image is adjusted through a brightness equalization mechanism (such as Gamma correction based on the average pixel value of the whole frame) to generate a style layer processing channel.

[0061] Based on semantic label features and object detection results, a semantic layer processing channel is constructed through semantic segmentation algorithm and feature alignment mechanism. The pixel layer processing channel, style layer processing channel and semantic layer processing channel are connected in series to form a multi-level desensitization processing channel.

[0062] Specifically, semantic label features (such as classification labels and confidence scores for targets like vehicles, pedestrians, and license plates) and object detection results are input into the semantic segmentation algorithm. Based on the prior localization information provided by the object detection results, pixel-level segmentation masks are generated to identify specific semantic category regions. At the same time, through feature alignment mechanisms (such as using optical flow fields or attention weight matrices to temporally align the segmentation results of adjacent frames), the masks of the same target in the video sequence are ensured to maintain spatial consistency in the temporal dimension, generating semantic layer processing channels.

[0063] The pixel layer processing channel (providing basic pixel modification), the style layer processing channel (providing appearance style transformation), and the semantic layer processing channel (providing precise semantic region positioning) are connected in sequence according to the processing flow to form a multi-level desensitization processing channel.

[0064] S3. The encrypted video data packets are subjected to feature extraction and semantic analysis through a multimodal fusion recognition model to obtain the recognition confidence and context information of sensitive elements, and a hierarchical judgment is performed to generate channel selection instructions.

[0065] The spatial distribution features and texture information of encrypted video data packets are extracted by the visual feature extraction branch of the multimodal fusion recognition model to generate a spatial feature set, while the temporal feature branch extracts the motion trajectory and inter-frame change information of encrypted video data packets to generate a temporal feature set.

[0066] Specifically, the visual feature extraction branch uses a convolutional neural network structure (e.g., ResNet-50 backbone) to process encrypted video data packets frame by frame. It extracts spatial distribution features describing the shape, contour, and relative position of objects through multi-layer convolution and pooling operations (e.g., generating feature maps of size 7x7x1024). At the same time, it captures the surface patterns and detailed information of the image, i.e., texture information (e.g., through Gabor filter bank responses or local binary pattern histograms), and combines the outputs into a spatial feature set.

[0067] The temporal feature branch of the multimodal fusion recognition model uses optical flow calculation algorithms (such as Farneback optical flow method) or three-dimensional convolutional networks (such as I3D) to analyze the pixel displacement between consecutive frames, extract motion trajectory information describing the target's moving path and speed (such as generating optical flow vector fields), calculate the difference between corresponding pixels in adjacent frames, obtain inter-frame change information (such as obtaining a binarized motion mask through frame difference method), and output a temporal feature set.

[0068] It should be noted that the multimodal fusion recognition model is trained by training a visual feature extraction branch (e.g., ResNet-50) on the ImageNet dataset, a temporal feature branch (e.g., I3D network) on the Kinetics dataset, and a semantic understanding branch (e.g., BERT-base) on the Wikipedia and BookCorpus corpora. These three branches are then combined with a randomly initialized feature fusion layer, a confidence calculation mechanism, and a hierarchical judgment logic. End-to-end fine-tuning is performed using a traffic video dataset containing sensitive target annotations and contextual information. All parameters are jointly optimized by minimizing the multi-task loss function (including target recognition cross-entropy loss and relationship prediction mean squared error loss), and iteratively updated using the Adam optimizer (with an example learning rate of 0.001) to obtain the trained multimodal fusion recognition model.

[0069] The semantic understanding branch of the multimodal fusion recognition model is used to extract contextual features and object correlation of encrypted video data packets, generating a semantic feature set.

[0070] Specifically, the semantic understanding branch of the multimodal fusion recognition model receives semantic label features and object detection results, encodes the text description in the semantic label features, and generates context features that capture the global semantics of the scene and the semantic relationships between elements;

[0071] Based on the bounding box coordinates and class probabilities of all detected targets in the object detection results, a graph structure is constructed (nodes are target instances, and edges are spatial overlaps or distance relationships). The message passing and aggregation between nodes are calculated through a graph convolutional network to generate quantified object association. The context features and object association are concatenated and dimensionality reduced to form a semantic feature set.

[0072] The feature fusion layer of the multimodal fusion recognition model fuses spatial feature sets, temporal feature sets, and semantic feature sets to generate a multimodal fusion feature tensor. Then, through the confidence calculation mechanism and context weight function, recognition confidence and context information are generated.

[0073] Specifically, the feature fusion layer of the multimodal fusion recognition model receives the spatial feature set, the temporal feature set of the temporal feature branch, and the semantic feature set of the semantic understanding branch. Through a fully connected layer, the features of different modalities are projected onto a unified dimension (e.g., projected onto a 1024-dimensional vector). A weighted fusion method based on an attention mechanism is used to weight and combine the spatial feature set, the temporal feature set, and the semantic feature set to generate a multimodal fusion feature tensor.

[0074] The multimodal fusion feature tensor is input into the confidence calculation mechanism, and a softmax classifier is used to calculate the class probability of each potential sensitive target (e.g., the output probability value example takes values ​​from 0 to 1) as the recognition confidence. The importance score of different context elements is calculated through the context weight function to generate context information.

[0075] Based on the recognition confidence level and contextual information, the desensitization level of sensitive targets is determined by the hierarchical judgment logic of the multimodal fusion recognition model, and the channel selection instruction is output.

[0076] Specifically, based on recognition confidence and contextual information, the desensitization level of sensitive targets is determined by the hierarchical judgment logic of the multimodal fusion recognition model. The hierarchical judgment logic receives the recognition confidence and contextual information output from the feature fusion layer, divides the recognition confidence into different intervals (for example, a confidence level below 0.3 is considered low sensitivity, 0.3 to 0.7 is considered medium sensitivity, and a confidence level above 0.7 is considered high sensitivity), and combines the importance scores of different elements in the contextual information (for example, elements with an importance score above 0.8 are given higher sensitivity weights) for weighted correction to obtain the corrected sensitivity level.

[0077] Each detection target is comprehensively evaluated, and contextual information is used to determine the scene type of the target and the relationship with other elements. Based on the corrected sensitivity level, the corresponding channel selection instruction is generated.

[0078] S4. Match the corresponding processing path in the multi-level desensitization channel according to the channel selection instruction, perform corresponding desensitization on different sensitive areas, and output desensitized video frames and sensitive area separation reports.

[0079] Select the appropriate desensitization processing channel according to the channel selection instruction, and assemble the corresponding desensitization processing channel and the multimodal fusion feature tensor to generate the data stream to be desensitized.

[0080] Specifically, based on the anonymization level information contained in the channel selection instruction (e.g., high sensitivity level corresponds to "pixel layer processing channel + semantic layer processing channel"), the corresponding processing path is selected from the multi-level anonymization processing channels according to the processing channel combination specified in the channel selection instruction (e.g., different combinations of pixel layer, style layer, or semantic layer processing channels); the coordinate information, feature vector, and associated metadata (e.g., timestamp and object identifier) ​​of the sensitive target are extracted from the multimodal fusion feature tensor; the selected processing path is bound with the extracted coordinate information and feature vector to form structured processing configuration parameters; and the original video frame data is integrated with the processing configuration parameters to generate the data stream to be anonymized.

[0081] Extract the coordinate information of the sensitive area from the data stream to be de-identified, perform the de-identification operation, generate the de-identified feature frame, verify the integrity of the de-identified area boundary through the topology consistency comparison algorithm, and output the corrected de-identified feature frame.

[0082] Specifically, based on the processing configuration parameters bound in the data stream to be desensitized, the corresponding adaptive desensitization strategy is selected (such as selecting the adaptive blur algorithm and edge-preserving filtering mechanism for the pixel layer processing channel, or the image style transfer algorithm and brightness equalization mechanism for the style layer processing channel), and the desensitization operation is performed (for example, applying Gaussian blur kernel size of 15x15 to the license plate area); the processed frame data is combined with the feature descriptor in the multimodal fusion feature tensor to generate the desensitized feature frame;

[0083] The geometric consistency between the desensitized region and the original sensitive region is verified by a topological consistency comparison algorithm (a boundary comparison method based on morphological dilation and erosion operations and contour detection). Boundary gaps or overflows are detected, and interpolation repair or edge clipping is performed on the detected inconsistent regions. The corrected desensitized feature frame is then output.

[0084] The frame sequence reconstruction process restores the corrected desensitized feature frame order, generates desensitized video frames, and aggregates the type, location, and processing intensity of sensitive areas to generate a sensitive area separation report.

[0085] Specifically, based on the timestamp information embedded in the corrected desensitized feature frames, the disordered corrected desensitized feature frames are sorted and reassembled according to the original time sequence; the reassembled frame sequence is decoded into a continuous video stream to generate desensitized video frames.

[0086] Simultaneously, the desensitization processing records of all sensitive regions are extracted from the metadata of each corrected desensitized feature frame, including sensitive region type (e.g., "license plate", "face"), location information and processing intensity parameters, and aggregated by frame sequence to generate a sensitive region separation report.

[0087] S5. Compare the structural features of the original video frames and the desensitized video frames, calculate the structural difference index, and generate structural consistency verification information.

[0088] Based on the image feature data of the original video frames and the desensitized video frames, edge distribution features, texture gradient features and motion trajectory features are obtained through feature extraction algorithms to obtain a multi-dimensional feature set.

[0089] Specifically, the Canny edge detection algorithm is applied to both the original and desensitized video frames to extract edge distribution features describing the shape of the object's contour; the Local Binary Pattern (LBP) operator is used to calculate texture gradient features (generating a 256-dimensional texture histogram), and combined with the Histogram of Oriented Gradients (HOG) to capture more refined texture structure information (the cell size is 8x8 pixels in the example); the Farneback optical flow algorithm is used to calculate the displacement vectors of corresponding points between consecutive frames to extract motion trajectory features describing the target's motion pattern (generating an optical flow vector field, with 3 pyramid layers in the example); the three types of features from the original and desensitized video frames are aligned and combined according to frame time sequence to form a multi-dimensional feature set.

[0090] The structural difference index is calculated using the structural similarity calculation formula and the optical flow consistency function. The structural difference index is then mapped to the structural preservation level. The structural difference index and the structural preservation level are jointly encapsulated to generate structural consistency verification information.

[0091] Specifically, the Structural Similarity Calculation Formula (SSIM) is used to calculate the similarity between the original video frame and the desensitized video frame in three components: brightness, contrast, and structure, generating a structural similarity value. The Euclidean distance between the corresponding motion trajectory features of the original video frame and the desensitized video frame is calculated using the Optical Flow Consistency Function, generating an optical flow consistency error value. The structural similarity value and the optical flow consistency error value are weighted and fused (with a weighting coefficient of 0.5 for each in the example) to generate a structural difference index. The structural difference index is mapped to a structural preservation level (e.g., a difference below 0.1 is "excellent", 0.1 to 0.3 is "good", and above 0.3 is "average"). The structural difference index and the structural preservation level are then encapsulated together to generate structural consistency verification information.

[0092] It should be noted that the expression for calculating the structural difference index is as follows:

[0093]

[0094] in, It is a structural difference index. This is the weighting coefficient for the spatial structure difference item, representing its proportion in the overall evaluation. It is a constant used to convert SSIM from similarity to difference. It is a structural similarity index. These are the original video frames. These are the anonymized video frames. These are the weighting coefficients for the temporal motion consistency error term. It is an optical flow consistency error.

[0095] S6. Unify the de-identified video frames, structural consistency verification information, and sensitive area separation reports into a de-identified data package.

[0096] Based on desensitized video frames, structural consistency verification information, and sensitive area separation reports, a corresponding relationship is established through a spatiotemporal label matching mechanism to generate an encapsulation index table, which is then mapped and integrated to generate a data fusion structure.

[0097] Specifically, the timestamps and spatial metadata embedded in the desensitized video frames are extracted. Simultaneously, the corresponding frame identifiers and structural difference indicators are obtained from the structural consistency verification information. Sensitive area processing records (including type, coordinates, and processing intensity) under the same timestamp are extracted from the sensitive area separation report. A spatiotemporal tag matching mechanism is used to precisely align the data according to the timestamps and spatial coordinates, generating a unique index entry for each match. All index entries are organized into a packaged index table. The original data from the desensitized video frames, structural consistency verification information, and sensitive area separation reports are binary mapped and concatenated based on the offset information in the packaged index table to generate a data fusion structure.

[0098] Based on the data fusion structure, a unique verification code is generated through hash encryption and timestamp identification mechanism. The unique verification code and the encapsulation index table are serialized and combined to generate a de-identified data packet.

[0099] Specifically, the SHA-256 hash algorithm is applied to the complete binary content of the data fusion structure to generate a hash value, and the current Coordinated Universal Time timestamp is obtained at the same time; the timestamp and the hash value are concatenated and then subjected to SHA-256 hash operation again to generate a unique check code;

[0100] The uniqueness check code and the encapsulation index table are combined using the Protocol Buffers serialization protocol in binary encoding. The encapsulation index table serves as the main data block, and the uniqueness check code serves as the header verification field. The data is then concatenated byte-by-byte according to the standard data packet format to generate a de-identified data packet.

[0101] This embodiment also provides an intelligent desensitization processing system for traffic videos, including: an acquisition module for acquiring raw video streams, encrypting and transmitting them to a trusted data space via a privacy protection gateway, and simultaneously attaching spatiotemporal metadata to generate encrypted video data packets; a construction module for establishing a multi-level desensitization channel based on the encrypted video data packets, including pixel-level processing, style-level processing, and semantic-level processing; an instruction module for performing feature extraction and semantic analysis on the encrypted video data packets using a multimodal fusion recognition model, obtaining the recognition confidence and context information of sensitive elements, performing hierarchical judgment, and generating channel selection instructions; a desensitization module for matching corresponding processing paths in the multi-level desensitization channels according to the channel selection instructions, performing corresponding desensitization on different sensitive areas, and outputting desensitized video frames and sensitive area separation reports; a verification module for comparing the structural features of the original video frames and the desensitized video frames, calculating the structural difference index, and generating structural consistency verification information; and an encapsulation module for uniformly encapsulating the desensitized video frames, structural consistency verification information, and sensitive area separation reports to form a desensitized data packet.

[0102] This embodiment also provides a computer device applicable to the intelligent desensitization processing method for traffic videos, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent desensitization processing method for traffic videos as proposed in the above embodiment.

[0103] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0104] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent desensitization processing method for traffic videos as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0105] In summary, this invention achieves refined and adaptive hierarchical processing of multimodal sensitive information by constructing a multi-level desensitization channel, thus solving the problems of existing technologies having a single desensitization strategy and being unable to balance privacy protection and data availability. Under the premise of thoroughly protecting privacy, it significantly enhances the value and application potential of desensitized data in high-precision traffic analysis.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent desensitization processing of traffic videos, characterized in that: include, The raw video stream is captured, encrypted and transmitted to a trusted data space through a privacy protection gateway, and spatiotemporal metadata is attached to generate encrypted video data packets. The specific steps for generating encrypted video data are as follows: The original video stream is acquired and embedded with six-dimensional spatiotemporal metadata generated by satellite positioning to form a spatiotemporal tensor data packet. The pixel spatial topology of the original video frame is used to construct a dynamic Riemannian manifold and perform nonlinear distortion encryption to obtain the encrypted spatiotemporal tensor data packet. The encrypted spatiotemporal tensor data packet is divided into frame units, each frame unit is digitally signed and a Merkle tree is constructed, the root value of the Merkle tree is generated, and the geometric topological consistency of each frame unit with the dynamic Riemannian manifold is verified through a zero-knowledge proof protocol to generate a verified encrypted data stream. The verified encrypted data stream is reassembled into a first chain and a second chain according to the parity of frame units, and a manifold connectivity topology certificate is attached to obtain an encrypted video data packet. Based on encrypted video data packets, a multi-level desensitization channel is established, which includes pixel layer processing, style layer processing, and semantic layer processing. The process of establishing a multi-level desensitization channel based on encrypted video data packets, including pixel-level processing, style-level processing, and semantic-level processing, involves the following steps: The frame sequence of encrypted video data packets is parsed to extract spatial feature data, pixel distribution features, time series features, semantic tag features, and object detection results; Based on spatial feature data and pixel distribution features, a pixel layer processing channel is constructed through an adaptive blur algorithm and an edge-preserving filtering mechanism. Based on spatial feature data and time series features, a style layer processing channel is constructed through an image style transfer algorithm and a brightness equalization mechanism. Based on semantic label features and object detection results, a semantic layer processing channel is constructed through semantic segmentation algorithm and feature alignment mechanism. The pixel layer processing channel, style layer processing channel and semantic layer processing channel are connected in series to form a multi-level desensitization processing channel. The encrypted video data packets are subjected to feature extraction and semantic analysis by a multimodal fusion recognition model to obtain the recognition confidence and context information of sensitive elements, and a hierarchical judgment is performed to generate channel selection instructions. According to the channel selection instruction, the corresponding processing path is matched in the multi-level desensitization channel to perform corresponding desensitization on different sensitive areas, and output desensitized video frames and sensitive area separation reports. The specific steps for outputting the desensitized video frames and sensitive region separation report are as follows. Select the corresponding desensitization processing channel according to the channel selection instruction, and assemble the corresponding desensitization processing channel and the multimodal fusion feature tensor to generate the data stream to be desensitized; Extract the coordinate information of the sensitive area from the data stream to be de-identified, perform the de-identification operation, generate the de-identified feature frame, verify the integrity of the de-identified area boundary through the topology consistency comparison algorithm, and output the corrected de-identified feature frame; The frame sequence reconstruction process restores the corrected desensitized feature frame order, generates desensitized video frames, and aggregates the type, location, and processing intensity of sensitive areas to generate a sensitive area separation report. The structural features of the original video frames and the desensitized video frames are compared, the structural difference index is calculated, and structural consistency verification information is generated. The de-identified video frames, structural consistency verification information, and sensitive area separation reports are uniformly packaged to form a de-identified data package.

2. The intelligent desensitization processing method for traffic videos as described in claim 1, characterized in that: The specific steps for generating the channel selection instruction are as follows: The spatial distribution features and texture information of encrypted video data packets are extracted by the visual feature extraction branch of the multimodal fusion recognition model to generate a spatial feature set, and the temporal feature branch is extracted by the motion trajectory and inter-frame change information of encrypted video data packets to generate a temporal feature set. The semantic understanding branch of the multimodal fusion recognition model is used to extract contextual features and object correlation of encrypted video data packets to generate a semantic feature set; The feature fusion layer of the multimodal fusion recognition model fuses spatial feature sets, temporal feature sets, and semantic feature sets to generate a multimodal fusion feature tensor. Then, through the confidence calculation mechanism and context weight function, recognition confidence and context information are generated. Based on recognition confidence and contextual information, the desensitization level of sensitive targets is determined by the hierarchical judgment logic of the multimodal fusion recognition model, and the channel selection instruction is output.

3. The intelligent desensitization processing method for traffic videos as described in claim 1, characterized in that: The specific steps for generating structural consistency verification information are as follows. Based on the image feature data of the original video frames and the desensitized video frames, edge distribution features, texture gradient features and motion trajectory features are obtained through feature extraction algorithms to obtain a multi-dimensional feature set. The structural difference index is calculated using the structural similarity calculation formula and the optical flow consistency function. The structural difference index is then mapped to the structural preservation level. The structural difference index and the structural preservation level are jointly encapsulated to generate structural consistency verification information.

4. The intelligent desensitization processing method for traffic videos as described in claim 1, characterized in that: The specific steps for forming the de-identified data packet are as follows: Based on desensitized video frames, structural consistency verification information, and sensitive area separation reports, a corresponding relationship is established through a spatiotemporal label matching mechanism to generate an encapsulation index table, which is then mapped and integrated to generate a data fusion structure. Based on the data fusion structure, a unique verification code is generated through hash encryption and timestamp identification mechanism. The unique verification code and the encapsulation index table are serialized and combined to generate a de-identified data packet.

5. A traffic video intelligent desensitization processing system, based on the traffic video intelligent desensitization processing method according to any one of claims 1 to 4, characterized in that: include, The acquisition module is used to acquire raw video streams, encrypt and transmit them to a trusted data space through a privacy protection gateway, and simultaneously attach spatiotemporal metadata to generate encrypted video data packets. The module is used to establish a multi-level desensitization channel based on encrypted video data packets, which includes pixel layer processing, style layer processing, and semantic layer processing. The instruction module is used to extract features and perform semantic analysis on encrypted video data packets through a multimodal fusion recognition model, obtain the recognition confidence and context information of sensitive elements, perform hierarchical judgment, and generate channel selection instructions. The desensitization module is used to match the corresponding processing path in the multi-level desensitization channel according to the channel selection instruction, perform corresponding desensitization on different sensitive areas, and output desensitized video frames and sensitive area separation reports; The verification module is used to compare the structural features of the original video frames with those of the desensitized video frames, calculate the structural difference index, and generate structural consistency verification information. The encapsulation module is used to encapsulate the de-identified video frames, structural consistency verification information, and sensitive area separation reports in a unified manner to form a de-identified data package.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent desensitization processing method for traffic videos as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent desensitization processing method for traffic videos as described in any one of claims 1 to 4.

Images