Intelligent video monitoring management method and system based on big data

Abstract

This invention discloses an intelligent video surveillance management method and system based on big data, relating to the field of video surveillance technology. The method includes using a VIBE model to generate SMPL-X model parameters to obtain the vertex coordinates of the head in a 3D human body mesh; using a DECA model to generate FLAME parameters to obtain the vertex coordinates of the face in a 3D face mesh; and fusing the vertex coordinates of the human and face meshes. By introducing a keypoint extraction method combining traditional image processing and deep learning, the accuracy of keypoint detection in occluded and low-light scenes is improved, enhancing the robustness of 3D human body modeling. Dynamic region masking and the YOLACT++ model are used to optimize the semantic feature extraction of moving regions. Through fixed-point coordinate fusion optimization, a high-precision 3D human body model is generated. Combined with encryption and multi-level LoP video output, it flexibly adapts to high-privacy and resource-constrained scenarios, improving the adaptability and practicality of the surveillance system.

Description

Technical Field

[0001] This invention relates to the field of video surveillance technology, and in particular to intelligent video surveillance management methods and systems based on big data. Background Technology

[0002] In recent years, with the rapid development of computer vision and artificial intelligence technologies, intelligent video surveillance systems have been widely used in public safety, urban management, and retail. However, existing technologies still have several shortcomings. Traditional human pose estimation and 3D modeling methods lack robustness in complex scenarios (such as occlusion, low light, or multi-person interaction), leading to decreased accuracy in key point detection and distortion in 3D mesh reconstruction. Secondly, existing methods lack efficient motion region segmentation and semantic feature extraction mechanisms when processing dynamic regions, making it difficult to achieve refined anonymization in high-privacy scenarios. Furthermore, the fusion process of 3D human body and facial meshes usually relies on complex parameter optimization, resulting in high computational overhead and difficulty in meeting the real-time requirements of edge devices. Although the application of data compression and encryption technologies has improved transmission efficiency, the balance between semantic information preservation and privacy protection still needs optimization. When generating semantically anonymized videos, existing monitoring systems struggle to flexibly adapt to different privacy levels and resource constraints, limiting their widespread adoption in diverse application scenarios. Existing intelligent video surveillance technologies have significant problems in terms of robustness, real-time performance, and the balance between privacy protection and semantic information preservation in complex scenarios. Our invention proposes a big data-based intelligent video surveillance management method that solves the accuracy and efficiency problems of 3D human body modeling in complex scenarios. It optimizes the semantic feature extraction and data compression and encryption processes for dynamic areas and achieves multi-level anonymized video output to adapt to different privacy and resource requirements. This invention belongs to the fields of computer vision, 3D modeling, and privacy protection technology, and combines deep learning, image processing, and data encryption technologies to provide an innovative solution for intelligent video surveillance. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a big data-based intelligent video surveillance management method and system, which solves the significant problems of existing intelligent video surveillance technologies in terms of robustness, real-time performance, and the balance between privacy protection and semantic information preservation in complex scenarios.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides an intelligent video surveillance management method based on big data, which includes generating standardized low-resolution images and human body bounding boxes by acquiring raw video streams and performing preprocessing.

[0007] Based on standardized images and bounding boxes, 2D key point coordinates and confidence scores of the human body and face are extracted through morphological thinning algorithms and MTCNN models. The average confidence score of facial key points is calculated, and facial key points and confidence scores are dynamically updated.

[0008] Based on human key points and confidence levels, the VIBE model is used to generate SMPL-X model parameters to obtain the vertex coordinates of the head of the 3D human body mesh. Based on dynamically updated facial key points and confidence levels, the DECA model is used to generate FLAME parameters to obtain the vertex coordinates of the 3D facial mesh. The vertex coordinates of the human body and facial mesh are fused, and the SMPL-X and FLAME parameters are compressed.

[0009] The Z-Standard decoding algorithm is used to decompress the dynamic region mask, and the 3D localization and semantic feature vector of the object are generated based on the dynamic region mask.

[0010] Generate semantic data compression and encryption data packets based on semantic feature vectors;

[0011] Decompress the encrypted data packet, fuse 3D model parameters and object semantic feature vectors, and generate anonymized videos with semantic annotations.

[0012] As a preferred embodiment of the big data-based intelligent video surveillance management method of the present invention, wherein: the step of acquiring the original video stream and preprocessing it to generate a standardized low-resolution image and human body bounding box refers to using Gaussian filtering on the original frame I t Denoising is performed to obtain a smooth frame I' t The current frame I is calculated using the inter-frame difference method. t Compared to the previous frame I t-1 If the sum of the pixel differences is less than a threshold k, it is marked as a static frame, only its timestamp is recorded, and subsequent processing is skipped. If the sum of the differences is greater than or equal to the threshold k, it is marked as a non-static frame, and I'... t Scaled to 512×512 resolution to generate a normalized image I” t Using the YOLOv5s model in normalized image I” t Detects a human body and outputs the bounding box coordinates and confidence score c. b Calculate the normalized image I of the current frame. t With the previous frame I” t-1 The pixel differences are used to generate a difference image. Then, a threshold A is applied to binarize the difference image to generate a binary mask E. t Regions with pixel differences greater than or equal to threshold A are marked as 1, representing moving regions, while regions with pixel differences less than threshold A are marked as 0, representing static regions.

[0013] As a preferred embodiment of the big data-based intelligent video surveillance management method of the present invention, the method involves: extracting 2D key point coordinates and confidence scores of the human body and face based on standardized images and bounding boxes using morphological thinning algorithms and the MTCNN model; calculating the average confidence score of facial key points; and dynamically updating the facial key points and confidence scores. The standardized image is then processed using a morphological thinning algorithm. t China B t For the corresponding region, extract the human skeleton image to represent the human centerline structure. Use Hough transform to detect straight line segments in the skeleton image to identify limb parts. Locate 25 key points through geometric constraints, calculate the confidence score for each key point, and generate the key point position x. t and confidence vector c t According to B t Trim the human body ROI region R t ;

[0014] For the trimmed R t 68 facial landmarks were extracted using the MTCNN model to generate facial landmarks. and confidence vector Calculate the average confidence score of facial key points like If the threshold Q is greater than or equal to the current frame's threshold, then the current frame's threshold is used directly. As output, input to the DECA model, if <Threshold Q, extract dynamic region mask E t For regions with a median value of 1, rerun MTCNN within the masked dynamic region to obtain new facial key points. and confidence level Use new As input to the DECA model, if new If the result is greater than or equal to the threshold Q, use the new result; otherwise, use the facial landmarks and confidence scores from the previous frame.

[0015] As a preferred embodiment of the intelligent video surveillance management method based on big data described in this invention, the following steps are performed: Based on human key points and confidence levels, SMPL-X model parameters are generated using the VIBE model to obtain the 3D human head mesh vertex coordinates; based on dynamically updated facial key points and confidence levels, FLAME parameters are generated using the DECA model to obtain the 3D face mesh vertex coordinates; the vertex coordinates of the human and face meshes are fused; and the SMPL-X and FLAME parameters are compressed. The VIBE model is then used based on human key points x... t and confidence level c t Generate the keypoint rotation parameters θ of the SMPL-X model. t Shape parameter β t Translation parameter T tThe keypoint parameters are input into the SMPL-X model, and the vertex positions of the 3D mesh are calculated based on the forward kinematics function, outputting the 3D human body mesh vertex U. t For rotation parameter θ t Perform time series optimization and calculate smoothing rotation parameters. Extracting fixed points v of the head region from the vertices of a 3D human body mesh. i Coordinates, based on the predefined head vertex index i of the SMPL-X model, generate a fixed point coordinate set, denoted as V. SMPL-X ;

[0016] Using the DECA model, based on 68 dynamically updated facial landmarks using MTCNN. and confidence level Generate FLAME model parameters, including facial pose. Facial shape Facial expression parameter k t The FLAME model outputs the 3D mesh of the face, specifically the Y-axis. t Update facial shape parameters Based on 68 facial key points output by the MTCNN model, a subset of 3D mesh vertices corresponding to the 68 key points is obtained using the predefined topology of the FLAME model. Through the key point-to-vertex index mapping provided by FLAME, core vertices of key facial regions are selected. The weighted average of the 3D coordinates of the core vertices is calculated to generate the facial vertex coordinates u. fixed ;

[0017] By using the predefined SMPL-X head vertex index and FLAME model face vertex mapping, a V... SMPL-X The correspondence between the head vertices and the FLAME mesh vertices is determined using the iterative nearest-point algorithm. The rigid transformation matrix of the FLAME mesh vertices is calculated, such that u... fixed Facial vertex and V SMPL-X The head vertices are aligned, and the aligned FLAME mesh vertices and SMPL-X head mesh vertices are weighted and fused to generate a new 3D mesh vertex V. new The merged head vertex V new The body mesh vertices are merged with those of SMPL-X, and the Laplacian smoothing algorithm is applied to optimize the mesh geometric continuity, generating a unified 3D human body model. Differential coding is applied to the SMPL-X and FLAME parameters to record inter-frame parameter changes, and the dynamic region mask E is used. t Flattened into a 1D binary sequence, the input is the Z-Standard compression algorithm. The binary sequence is compressed using dictionary compression and entropy coding mechanisms to generate compressed mask data. Huffman coding is then used to perform secondary compression on the differencing parameters and the Z-Standard compressed mask to generate smaller anonymized data packets.

[0018] As a preferred embodiment of the big data-based intelligent video surveillance management method of the present invention, the following steps are included: Decompressing the dynamic region mask using the Z-Standard decoding algorithm, generating 3D positioning and semantic feature vectors of objects based on the dynamic region mask, and using the Z-Standard decoding algorithm to decompress the dynamic region mask, recovering the binary mask, applying the YOLACT++ model only to regions with a dynamic region mask value of 1 on the standardized image, generating object masks, object categories, and confidence levels, filtering objects with confidence levels greater than or equal to a threshold G, obtaining the center point coordinates of the object mask, and using the Monodepth2 model on the standardized image I. t Perform depth calculations to generate a depth map D. t Extract the depth value of the center point of the human body, and the Y-axis of the human body 3D mesh vertex. t Using a coordinate system, the center point of the object is mapped from 2D to 3D space, generating the object's 3D positioning coordinates. A PointNet++ model is then used to perform semantic analysis on the pixel point cloud of the object mask, generating the object's semantic feature vector. This is achieved from the vertices U of the human body 3D mesh. t Extract the vertices of the hand and combine them with the 3D mesh vertices of the face (Y). t Generate local 3D scene features.

[0019] As a preferred embodiment of the big data-based intelligent video surveillance management method of the present invention, wherein: the generation of semantic data compression and encryption data packets based on semantic feature vectors refers to using the MeshLab algorithm to process new 3D mesh vertices V new Vertex simplification is performed to generate compressed mesh data. The object mask is flattened into a 1D bit string, and a compressed object mask is generated using the Brotli compression algorithm. The object's 3D positioning coordinates, local 3D scene features, and semantic feature vectors are quantized into 16-bit floating-point numbers to form a semantic parameter vector. The semantic parameter vector, compressed object mask, and object category label are packaged into Protobuf format data. The Protobuf data is encrypted using the XChaCha20 encryption algorithm to generate an encrypted data packet. The encrypted data packet is transmitted to the cloud server via a 5G network using the gRPC protocol.

[0020] As a preferred embodiment of the big data-based intelligent video surveillance management method of the present invention, the following steps are performed: Decompressing the encrypted data packet, fusing 3D model parameters and semantic feature vectors to generate anonymized video data with semantic annotations, decrypting the encrypted data packet using the XChaCha20 decryption algorithm to recover the semantic parameter vector, compressed object mask, object category label, object 3D positioning coordinates, and semantic feature vector, restoring the object mask using the Brotli decoding algorithm, embedding a facial mesh into the human body mesh based on a unified 3D human body model, mapping the object 3D positioning coordinates and semantic feature vectors to the 3D scene to generate a semantic scene representation, and using the Unreal Engine renderer to... By combining 3D human models, object masks, object category labels, and semantic features, a LoP4 level anonymized video is generated. The video is labeled with object categories and semantic descriptions. Based on the 3D scene data of LoP4, a lower-level LoP video is generated. LoP1 extracts the joint lines of the 3D human model and renders them as 2D skeleton lines. LoP2 overlays the object mask on the LoP1 to display the object outline. LoP3 retains the 3D human and object mesh, but removes textures and semantic annotations. The output level is selected according to the monitoring scene requirements: high privacy scenes output LoP1, medium privacy scenes output LoP2 or LoP3, and real-time and storage-constrained scenes prioritize LoP1 or LoP2.

[0021] Secondly, the present invention provides an intelligent video surveillance management system based on big data, including a video preprocessing and human detection module, used to acquire raw video streams and perform preprocessing to generate standardized low-resolution images and human bounding boxes.

[0022] The key point extraction and dynamic update module is used to extract 2D key points and confidence scores of the human body and face based on standardized images and bounding boxes.

[0023] The 3D modeling and parameter compression module is used to generate SMPL-X and FLAME parameters based on 2D key points, output 3D human body and face mesh vertices, and compress parameters;

[0024] The object localization and semantic analysis module is used to generate 3D localization and semantic features of objects based on dynamic region masks and 3D mesh vertices;

[0025] The data compression and encrypted transmission module is used to compress 3D mesh vertices and semantic data and encrypt them for transmission to the cloud.

[0026] 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 video surveillance management method based on big data as described in the first aspect of the present invention.

[0027] 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 video surveillance management method based on big data as described in the first aspect of the present invention.

[0028] The beneficial effects of this invention are as follows: By introducing a key point extraction method that combines traditional image processing with deep learning, the accuracy of key point detection in occluded and low-light scenes is effectively improved, enhancing the robustness of 3D human body modeling; the semantic feature extraction of moving regions is optimized using dynamic region masking and the YOLACT++ model, and combined with Z-Standard and Brotli compression algorithms, computational and storage overhead is significantly reduced, improving the real-time processing capability of edge devices; through fixed-point coordinate fusion optimization of SMPL-X and FLAME models, a high-precision unified 3D human body model is generated, and combined with XChaCha20 encryption and multi-level LoP video output, it flexibly adapts to high privacy and resource-constrained scenarios, ensuring a balance between semantic information preservation and privacy protection, significantly improving the adaptability and practicality of the monitoring system, and providing an efficient and reliable solution for intelligent video surveillance in complex scenarios. Attached Figure Description

[0029] 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.

[0030] Figure 1 This is a flowchart of the intelligent video surveillance management method based on big data in Example 1.

[0031] Figure 2 This is a schematic diagram of the structure of the big data-based intelligent video surveillance management system in Example 1.

[0032] Figure 3 This is a flowchart illustrating the generation of multi-level anonymous video in the big data-based intelligent video surveillance management method described in Example 1. Detailed Implementation

[0033] 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.

[0034] 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.

[0035] 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.

[0036] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides an intelligent video surveillance management method based on big data, including the following steps:

[0037] S1. By acquiring the raw video stream and preprocessing it, standardized low-resolution images and human bounding boxes are generated.

[0038] Based on standardized images and bounding boxes, the OpenPose model and MTCNN model are used to extract the coordinates and confidence of 2D key points of the human body and face, calculate the average confidence of facial key points, and dynamically update facial key points and confidence.

[0039] Specifically, by acquiring the raw video stream and preprocessing it, standardized low-resolution images and human bounding boxes are generated. This involves applying a Gaussian filter (standard deviation 1.5) to the raw frame I. t (Acquire raw video frames from the high-definition camera on the edge device) t Denoising is performed to obtain a smooth frame I' t To reduce noise interference with subsequent detection, the current frame I is calculated using the inter-frame difference method. t Compared to the previous frame I t-1 If the sum of the pixel differences is less than a threshold k (set through empirical parameter tuning), it is marked as a static frame, only its timestamp is recorded, and subsequent processing is skipped. If the sum of the differences is greater than or equal to the threshold k, it is marked as a non-static frame, and I' t Scaled to 512×512 resolution to generate a normalized image I” t The scaling ratio (1080 / 512) was recorded for coordinate reconstruction, and the YOLOv5s model was used in the normalized image I. t Detect the human body and output the human body bounding box coordinates B. t = {x, y, w, h} (coordinates (x, y), width w, height h) and confidence level c b(b is the identifier of the bounding box), locates the rough area (rectangle) of the human body in the image, without extracting specific key points, only providing location information;

[0040] Calculate the normalized image I of the current frame. t ∈R 512×512 With the previous frame I” t-1 The pixel differences are used to generate a difference image. Then, a threshold A (an empirical value obtained through empirical parameter tuning, such as 0.1) is applied to binarize the difference image, generating a 512×512 binary mask E. t Regions with pixel differences greater than or equal to threshold A are marked as 1, representing moving regions, while regions with pixel differences less than threshold A are marked as 0, representing static regions.

[0041] By applying Gaussian filtering (standard deviation 1.5) to the original video frames for denoising, smooth frames are generated, effectively reducing noise interference with subsequent detection and improving image processing stability in complex scenes (such as low light or occlusion). The use of inter-frame differencing combined with empirically tuned thresholds accurately distinguishes between static and non-static frames, significantly reducing computational load and optimizing the real-time performance of edge devices. The YOLOv5s model efficiently detects human bounding boxes, outputting only coarse location information, reducing the computational overhead of keypoint extraction. Simultaneously, a 512×512 binary mask is generated through inter-frame differencing and threshold A, accurately segmenting dynamic regions and providing a reliable foundation for subsequent semantic feature extraction. These improvements not only enhance the system's robustness and real-time processing capabilities in complex scenes but also lay the foundation for efficient 3D modeling and privacy protection, flexibly adapting to diverse monitoring needs.

[0042] Furthermore, based on standardized images and bounding boxes, 2D keypoint coordinates and confidence scores of the human body and face are extracted using morphological thinning algorithms and the MTCNN model. The average confidence score of facial keypoints is calculated, and the facial keypoints and confidence scores are dynamically updated. The standardized images are then processed using morphological thinning algorithms. t China B tFor the corresponding regions, human skeleton images are extracted to represent the human centerline structure. Hough transform is used to detect straight line segments in the skeleton image to identify limb parts. Based on the COCO keypoint dataset standards (including nose, left and right eyes, left and right ears, left and right shoulders, left and right elbows, left and right wrists, left and right hips, left and right knees, left and right ankles, neck, chest, and pelvis), 25 keypoints are located through geometric constraints. For example, head keypoints (nose, eyes, ears) are located by matching high-intensity regions using a predefined head template in the top 1 / 5 of the bounding box. Shoulders and neck are located by the endpoints of horizontal skeleton line segments in the upper 1 / 3 of the bounding box and the head to the bounding box. (Midpoint of the heart-connection line is determined), calculate the confidence score for each keypoint (extract the gradient value of each keypoint location from the edge image generated by Canny edge detection and normalize it to the range of 0 to 1, reflecting the reliability of the edge strength, as the first component of the confidence score), calculate the length of the skeleton line segment from the keypoint to its neighboring keypoints (according to the COCO keypoint topology definition, such as from the shoulder to the elbow) and compare it with the expected length based on human proportions. If the length is close to the expected length, a higher connectivity score (range of 0 to 1) is assigned as the second component of the confidence score, and check whether the keypoint is located in the binary mask E. t The dynamic region is defined, and if it is within the dynamic region, a fixed bonus of 0.3 is assigned; otherwise, it is 0. This serves as the third component of the confidence score. These three components (edge ​​strength, connectivity score, and dynamic region bonus) are added together with weights of 0.4, 0.3, and 0.3 respectively to obtain the confidence value for each keypoint, ranging from 0 to 1, ensuring that the reliability of the keypoint and its adaptability to dynamic scenes are reflected. This generates the keypoint location x. t and confidence vector c t ;

[0043] According to B t Trim the human body ROI region (from bounding box B) t Obtain coordinates and dimensions: top-left corner coordinates (x, y), width w, height h, in the normalized image I” t Extract the rectangular region: from (x,y) to (x+w,y+h). If the bounding box exceeds the image area, fill the edge pixels (using zero padding). Output the cropped ROI region. Where H ∈ ≈h, W ∈ ≈w, where ∈ is the index of the clipping region;

[0044] For the trimmed R t The MTCNN model was used to extract 68 facial keypoints (the MTCNN model follows the Dlib 68-point facial keypoint standard, covering facial contours, eyes, nose, and mouth feature points), generating facial keypoints and confidence vectors:

[0045]

[0046] in, For facial key points, Confidence level;

[0047] Calculate the average confidence score of facial key points (Obtain the average of 68 facial key points and confidence vectors from the MTCNN output). If the threshold Q is greater than or equal to the current frame's threshold, then the current frame's threshold is used directly. As output, input to the DECA model, if <Threshold Q, extract dynamic region mask E t The region with a median value of 1 (representing the motion region) is within the mask dynamic region (E). t Rerun MTCNN within the specified time (when the value is 1) to obtain new facial key points. and confidence level Use new As input to the DECA model, if new If the result is greater than or equal to the threshold Q, use the new result; otherwise, use the facial landmarks and confidence scores from the previous frame and dynamically update the facial landmarks and confidence scores to avoid detection failures caused by occlusion or low resolution.

[0048] By employing morphological thinning algorithms and Hough transform, the human skeleton is extracted from standardized images and bounding boxes, and 25 COCO standard keypoints are located. Combined with geometric constraints and edge strength analysis, the accuracy and stability of keypoint detection in occluded and low-resolution scenes are significantly improved. The MTCNN model is used to dynamically update 68 facial keypoints. Through confidence weighting and dynamic region masking optimization, the reliability of facial keypoints in complex dynamic scenes is ensured, overcoming detection failures caused by occlusion or low light. Keypoint confidence calculation integrates edge strength, skeleton connectivity, and dynamic region weights, effectively improving the robustness and adaptability of keypoint localization while reducing dependence on deep learning models and computational complexity. This not only enhances the accuracy and efficiency of keypoint extraction before 3D modeling but also provides high-quality input for subsequent parameter generation of SMPL-X and FLAME models, significantly improving the practicality and performance of intelligent video surveillance systems in diverse scenarios.

[0049] S2. Based on human body key points and confidence, use the VIBE model to generate SMPL-X model parameters and obtain the 3D human body mesh head vertex coordinates. Based on dynamically updated facial key points and confidence, use the DECA model to generate FLAME parameters and obtain the 3D facial mesh vertex coordinates. Merge the human body and facial mesh vertex coordinates and compress the SMPL-X and FLAME parameters.

[0050] The Z-Standard decoding algorithm is used to decompress the dynamic region mask, and the 3D localization and semantic feature vector of the object are generated based on the dynamic region mask.

[0051] Specifically, based on human keypoints and confidence levels, the VIBE model is used to generate SMPL-X model parameters to obtain the 3D human head vertex coordinates. Based on dynamically updated facial keypoints and confidence levels, the DECA model is used to generate FLAME parameters to obtain the 3D face core vertex coordinates. The vertex coordinates of the human and face meshes are fused, and the SMPL-X and FLAME parameters are compressed. The VIBE model (pre-trained on the Human3.6M and 3DPW datasets) is used based on human keypoints x... t and confidence level c t Generate the keypoint rotation parameters θ of the SMPL-X model. t Shape parameter β t Translation parameter T t The keypoint parameters are input into the SMPL-X model, and the vertex positions of the 3D mesh are calculated based on the forward kinematics function. (Based on the predefined topology of the SMPL-X model, a mapping table is obtained between 25 keypoints (such as nose, left and right eyes, left and right ears, left and right shoulders, left and right elbows, left and right wrists, left and right knees, and left and right ankles) from the COCO Keypoint dataset and 52 joints (including 22 for the body, 20 for the hands, and 10 for the face) in the SMPL-X model. This table is provided by the SMPL-X model and directly maps each keypoint to the corresponding joint or a combination of adjacent joints. For example, the left shoulder keypoint in COCO is mapped to the left shoulder joint in SMPL-X, and the left elbow keypoint is mapped to the left elbow joint.) The rotation parameters (local joint angles represented by quaternions) of the 25 keypoints output by the VIBE model are also considered. This mapping table assigns the rotation parameters of each keypoint to the corresponding SMPL-X joint. For joints not directly corresponding to keypoints (such as fine joints in the hand or face), an interpolation method (based on the rotation parameters of neighboring keypoints and the joint hierarchy of the SMPL-X model) is used to estimate the rotation parameters, generating a complete set of rotation parameters for 52 joints. The joint rotation parameters (quaternions) are converted into rotation matrices using the forward kinematics function of the SMPL-X model. Combined with shape and translation parameters, the vertex positions of the 3D human body mesh are calculated, ensuring the generated mesh matches the input keypoints. The output is the 3D human body mesh vertex U. t This represents the geometric structure of the human body in 3D space, using the formula:

[0052] U t =SMPL-X(θ) t ,β t ,T t ),

[0053] With rotation parameter θ t Time series optimization is performed, and smoothing rotation parameters are calculated based on keypoint confidence weights to ensure motion continuity. The formula is:

[0054]

[0055] in, Here, θ represents the optimized rotation parameter for the j-th keypoint, F is the time window set based on the continuity of actions in the monitoring scene, j = 25 is the number of keypoints in the SMPL-X model (corresponding to the 25 keypoints in COCO), defined by the SMPL-X model. key,j (f) represents the rotation parameters of the j-th keypoint in the f-th frame, and c j (f) represents the confidence level of the j-th keypoint, w m (f) represents the dynamic region mask weights for the f-th frame, generated based on the sparse optical flow algorithm. If keypoint j is located in the dynamic region, then w m =1, otherwise w m =0;

[0056] Extracting fixed points v of the head region from the vertices of a 3D human body mesh. i Coordinates (e.g., top of head, chin, ear), based on the predefined head vertex index i of the SMPL-X model, generate a fixed point coordinate set, denoted as V. SMPL-X ={v i |v i =(x i ,y i ,z i i = 1, ..., N head}, where N head This represents the total number of fixed point coordinates, which will be used for subsequent fusion optimization.

[0057] The DECA model (pre-trained on the VGGFace2 and CelebA datasets) was used, based on 68 facial keypoints dynamically updated by MTCNN. and confidence level Generate FLAME model parameters, including facial pose. Facial shape Facial expression parameter k t The FLAME model outputs the 3D mesh of the face, specifically the Y-axis. t (facial pose parameters) Shape parameters Facial expression parameter k t Input a FLAME model, and calculate the 3D mesh vertex positions using the model's geometric transformation function (a predefined facial template), representing the geometric structure of the face in 3D space. Formula:

[0058]

[0059] Update facial shape parameters To ensure time continuity, the formula is:

[0060]

[0061] in, The facial shape parameters are optimized from the previous frame, derived from the previous iteration (initial values ​​are the DECA model output), r t Average confidence level of facial key points e is the base of the natural logarithm (approximately 2.718), λ is the forgetting factor, obtained through experimental parameter tuning, and u is the frame index within the time window;

[0062] Based on 68 facial keypoints output by the MTCNN model, a subset of 3D mesh vertices corresponding to the 68 keypoints is obtained using the predefined topology of the FLAME model. Through the keypoint-to-vertex index mapping provided by FLAME (the mapping relationship between keypoints and FLAME mesh vertices is provided by the FLAME model, directly specifying the vertex index corresponding to each keypoint), core vertices of key facial regions (such as the tip of the nose, corners of the mouth, and corners of the eyes) are selected. The weighted average of the 3D coordinates of the core vertices is calculated, with the weight being the corresponding MTCNN keypoint confidence level. Generate face vertex coordinates u fixed =(x i ,y i ,z i );

[0063] By using the predefined SMPL-X head vertex index and FLAME model face vertex mapping, a V... SMPL-X The correspondence between the head vertices and the FLAME mesh vertices is determined using the Iterative Closest Point (ICP) algorithm. The rigid transformation matrix H = [R|t] of the FLAME mesh vertices is calculated, where R is the rotation matrix and t is the translation vector, such that u... fixed Facial vertex and V SMPL-X Head vertex alignment based on average confidence of facial key points in MTCNN and SMPL-X head keypoint confidence c t The fusion weight α is calculated by weighted averaging based on confidence level. The aligned FLAME mesh vertices and SMPL-X head mesh vertices are then weighted and fused to generate a new 3D mesh vertex V. new The merged head vertex V new By merging the body mesh vertices with SMPL-X and applying the Laplacian smoothing algorithm to optimize the mesh geometric continuity, a unified 3D human body model is generated, ensuring seamless connection between the head and body meshes.

[0064] V new =α·H(u fixed )+(1-α)·V SMPL-X ,

[0065] For SMPL-X parameters [θ' t ,β t ,T t ] and FLAME parameters Differential coding is applied to record inter-frame parameter changes (reducing redundancy), and the dynamic region mask E is used. t Flattened into a 1D binary sequence, input to the Z-Standard compression algorithm (configured with a high compression level, such as level=9), the binary sequence is compressed using dictionary compression and entropy coding mechanisms to generate compressed mask data. Huffman coding is then used to perform secondary compression on the differencing parameters and the Z-Standard compressed mask to generate a smaller anonymized data packet (<500 bytes).

[0066] To address the shortcomings of existing intelligent video surveillance technologies, such as insufficient 3D modeling accuracy, high computational overhead, and poor balance between privacy protection and semantic information preservation in complex scenes, this paper combines the VIBE and DECA models. Based on 25 COCO standard keypoints and 68 dynamically updated facial keypoints, SMPL-X and FLAME model parameters are generated respectively, efficiently constructing high-precision 3D human and facial meshes, significantly improving modeling robustness in occluded and dynamic scenes. A confidence-weighted time-series optimization method is used to smooth keypoint rotation parameters, ensuring motion continuity and reducing mesh distortion in complex scenes. The ICP algorithm and confidence-weighted averaging are used to fuse SMPL-X head vertices and FLAME facial vertices to generate a unified 3D human model, achieving seamless connection between head and body meshes and optimizing geometric consistency. In addition, differential coding combined with Z-Standard and Huffman compression algorithms is used to compress SMPL-X and FLAME parameters and dynamic region masks into bytes, significantly reducing storage and transmission overhead, adapting to the real-time requirements of edge devices, and significantly enhancing 3D modeling accuracy, computational efficiency, and privacy protection capabilities.

[0067] Furthermore, the process of using the Z-Standard decoding algorithm to decompress the dynamic region mask and generate 3D localization and semantic feature vectors of objects based on the dynamic region mask involves using the Z-Standard decoding algorithm to decompress the dynamic region mask, recovering a 512×512 binary mask to represent the motion region, and applying the YOLACT++ model only to regions with a dynamic region mask value of 1 on the standardized image to generate an object mask (512×512 binary matrix), object category (e.g., "knife" or "phone"), and confidence score. Objects with confidence scores greater than or equal to a threshold G (obtained through empirical parameter tuning based on object detection experiments on the COCO dataset and surveillance scene datasets (e.g., SIAT) are selected by statistically analyzing the confidence score distribution of the YOLACT++ model output and choosing a threshold G that effectively distinguishes between correct and false detections). The center point coordinates of the object mask are obtained, and the Monodepth2 model is used to process the standardized image I. t Perform depth calculations to generate a depth map D. t Extract the depth value of the center point of the human body, and the Y-axis of the human body 3D mesh vertex. t The coordinate system maps the object's center point from 2D (pixel coordinates + depth) to 3D space, generating the object's 3D positioning coordinates. The PointNet++ model is used to perform semantic analysis on the pixel point cloud of the object mask (extracting pixel coordinates (x, y) with a value of 1 from the object mask, combining this with the depth value of the corresponding pixel in the depth map Dt to generate point cloud data (each point contains x, y, z coordinates). This point cloud data is then input into the PointNet++ model, which performs semantic segmentation and feature extraction based on pre-trained weights (trained on the ShapeNet dataset), generating a 128-dimensional semantic feature vector of the object. This vector is then used to analyze the object's 3D mesh vertices U... t Extract approximately 500 vertices from the hand and combine them with the Y-axis of the 3D facial mesh. t Generate local 3D scene features (64 dimensions) to enhance scene semantics.

[0068] By efficiently decompressing binary dynamic region masks using the Z-Standard decoding algorithm, and applying the YOLACT++ model only to moving regions for object detection and segmentation, combined with Monodepth2 depth estimation and PointNet++ point cloud analysis, the method accurately generates 3D object positioning coordinates and 128-dimensional semantic feature vectors, significantly improving the accuracy and efficiency of object semantic extraction in complex dynamic scenes. Furthermore, by extracting hand and facial vertices from a unified 3D human mesh to generate 64-dimensional local scene features, the method further enhances the richness of scene semantics while reducing the computational overhead of non-dynamic regions. This approach effectively overcomes the lack of robustness in semantic information extraction in occluded or multi-person scenes using traditional techniques. Simultaneously, by focusing on dynamic regions, it optimizes computational resource utilization, adapts to the real-time requirements of edge devices, and provides high-quality input for subsequent generation of semantically anonymized videos, balancing privacy protection and semantic information.

[0069] S3. Generate semantic data compression and encryption data packets based on semantic feature vectors;

[0070] Decompress the encrypted data packet, fuse 3D model parameters and object semantic feature vectors, and generate anonymized videos with semantic annotations;

[0071] Specifically, the generation of semantic data compression and encryption data packets based on semantic feature vectors refers to using the MeshLab algorithm to process new 3D mesh vertices V. new Vertex simplification (reduced to approximately 50%) is performed to generate compressed mesh data. The object mask is flattened into a 1D bit string, and the Brotli compression algorithm is used to generate a compressed object mask with a compression ratio of approximately 1:1500. The object's 3D positioning coordinates (3D), local 3D scene features, and semantic feature vector (128D) are quantized into 16-bit floating-point numbers to form a semantic parameter vector. The semantic parameter vector, compressed object mask, and object category label are packaged into Protobuf format data. The Protobuf data is encrypted using the XChaCha20 encryption algorithm to generate an encrypted data packet. The encrypted data packet is transmitted to the cloud server via a 5G network using the gRPC protocol.

[0072] By simplifying 3D mesh vertices using the MeshLab algorithm and combining it with the Brotli compression algorithm to process object masks, the overhead of data storage and transmission is significantly reduced, optimizing the resource utilization of edge devices. 16-bit floating-point quantification of object 3D positioning coordinates, local scene features, and 128-dimensional semantic feature vectors is employed to generate efficient semantic parameter vectors, which are then packaged in Protobuf format and encrypted using the XChaCha20 encryption algorithm to generate encrypted data packets smaller than 500 bytes, ensuring high data security and privacy protection. Transmission to the cloud server via the gRPC protocol and 5G network further enhances the real-time performance and reliability of data transmission. This effectively solves the problems of data redundancy and privacy leakage in complex scenarios using traditional technologies, providing an efficient and secure data processing and transmission solution for intelligent video surveillance systems and significantly enhancing their applicability in diverse scenarios.

[0073] Furthermore, the decompressed encrypted data packet is fused with 3D model parameters and semantic feature vectors to generate an anonymized video index with semantic annotations. The encrypted data packet is then decrypted using the XChaCha20 decryption algorithm to recover the semantic parameter vector, compressed object mask, object category label, object 3D positioning coordinates, and semantic feature vector. The Brotli decoding algorithm is used to restore the 512×512 object mask. Based on a unified 3D human body model, a facial mesh is embedded into the human body mesh. The object 3D positioning coordinates and semantic feature vectors are mapped to the 3D scene to generate a semantic scene representation (including human and object positions). Using the Unreal Engine renderer, a 512×512×3 LoP is generated by combining the 3D human body model, object mask, object category label, and semantic features. Level 4 anonymity video, with object categories (e.g., "knife") and semantic descriptions labeled in the video. Based on 3D scene data of LoP4, it is downgraded to generate a lower-level LoP video. LoP1 (skeleton) extracts the joint lines of the 3D human model and renders them as 2D skeleton lines. LoP2 (skeleton + mask) overlays object masks on top of LoP1 to display the object outline. LoP3 (mesh) retains the 3D human and object mesh, but removes textures and semantic annotations. The output level is selected according to the monitoring scene requirements: high privacy scenes (such as public areas) output LoP1, medium privacy scenes (such as commercial retail) output LoP2 or LoP3, and real-time and storage-constrained scenes prioritize LoP1 or LoP2 to reduce rendering and storage overhead and meet different privacy and resource requirements.

[0074] The system efficiently recovers semantic parameter vectors and object masks through XChaCha20 decryption and Brotli decoding algorithms. Combined with a unified 3D human body model, facial meshes are seamlessly embedded into human body meshes, and object 3D positioning coordinates and semantic feature vectors are accurately mapped to the 3D scene to generate a semantically rich scene representation. Unreal Engine renderer is used to generate LoP4 level anonymized videos, which are combined with object categories and semantic descriptions. At the same time, LoP1 (skeleton), LoP2 (skeleton + mask), and LoP3 (mesh) videos are generated through downgrading, which can flexibly adapt to the needs of high privacy scenes (such as public areas), medium privacy scenes (such as commercial retail), and real-time / storage-constrained scenes. This multi-level anonymization output effectively balances privacy protection and semantic information preservation, significantly reduces rendering and storage overhead, and improves the robustness and adaptability of the system in complex scenes.

[0075] This embodiment also provides an intelligent video surveillance management system based on big data, including: a video preprocessing and human detection module, used to acquire raw video streams and perform preprocessing to generate standardized low-resolution images and human bounding boxes;

[0076] The key point extraction and dynamic update module is used to extract 2D key points and confidence scores of the human body and face based on standardized images and bounding boxes.

[0077] The 3D modeling and parameter compression module is used to generate SMPL-X and FLAME parameters based on 2D key points, output 3D human body and face mesh vertices, and compress parameters;

[0078] The object localization and semantic analysis module is used to generate 3D localization and semantic features of objects based on dynamic region masks and 3D mesh vertices;

[0079] The data compression and encrypted transmission module is used to compress 3D mesh vertices and semantic data and encrypt them for transmission to the cloud.

[0080] This embodiment also provides a computer device applicable to the intelligent video surveillance management method based on big data, 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 realize the intelligent video surveillance management method based on big data as proposed in the above embodiment.

[0081] 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.

[0082] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the intelligent video surveillance management method and system based on big data 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.

[0083] In summary, this invention effectively improves the accuracy of keypoint detection in occluded and low-light scenes and enhances the robustness of 3D human body modeling by introducing a keypoint extraction method that combines traditional image processing with deep learning. It optimizes the semantic feature extraction of moving regions using dynamic region masking and the YOLACT++ model, and significantly reduces computational and storage overhead by combining Z-Standard and Brotli compression algorithms, thereby improving the real-time processing capabilities of edge devices. Through fixed-point coordinate fusion optimization using SMPL-X and FLAME models, a high-precision unified 3D human body model is generated. Combined with XChaCha20 encryption and multi-level LoP video output, it flexibly adapts to high-privacy and resource-constrained scenarios, ensuring a balance between semantic information preservation and privacy protection. This significantly improves the adaptability and practicality of the monitoring system, providing an efficient and reliable solution for intelligent video surveillance in complex scenarios.

[0084] 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 big data-based intelligent video surveillance management method, characterized by: include, By acquiring the raw video stream and preprocessing it, standardized low-resolution images and human bounding boxes are generated. Based on standardized low-resolution images and human bounding boxes, 2D key point coordinates and their confidence scores of the human body and face are extracted through morphological thinning algorithms and MTCNN models. The average confidence score of facial key points is calculated, and facial key points and confidence scores are dynamically updated. Based on human key points and confidence scores, the VIBE model is used to generate SMPL-X model parameters to obtain the head vertex coordinates of the 3D human body mesh. The facial key points and confidence scores are dynamically updated. The DECA model is used to generate FLAME parameters to obtain the vertex coordinates of the 3D face mesh. The head vertex coordinates of the 3D human body mesh and the vertex coordinates of the 3D face mesh are fused, and the SMPL-X and FLAME parameters are compressed. The ZStandard decoding algorithm is used to decompress the dynamic region mask, and the 3D localization and semantic feature vector of the object are generated based on the dynamic region mask. Generate semantic data compression and encryption data packets based on semantic feature vectors; Decompress the encrypted data packet, fuse 3D model parameters and object semantic feature vectors, and generate anonymized videos with semantic annotations; The decompressed encrypted data packet is fused with 3D model parameters and semantic feature vectors to generate an anonymized video representation with semantic annotations. The encrypted data packet is then decrypted using the XChaCha20 decryption algorithm to recover the semantic parameter vectors, compressed object mask, object category labels, object 3D location coordinates, and semantic feature vectors. The object mask is then restored using the Brotli decoding algorithm. Based on a unified 3D human body model, a facial mesh is embedded into the human body mesh. The object 3D location coordinates and semantic feature vectors are mapped to the 3D scene to generate a semantic scene representation. This representation is then processed using Unreal Engine. The Engine renderer combines 3D human models, object masks, object category labels, and semantic features to generate LoP4 level anonymized videos. The videos are labeled with object categories and semantic descriptions. Based on the 3D scene data of LoP4, lower-level LoP videos are generated. LoP1 extracts the joint lines of the 3D human model and renders them as 2D skeleton lines. LoP2 overlays object masks on top of LoP1 to display the object outlines. LoP3 retains the 3D human and object meshes but removes textures and semantic annotations. The output level is selected according to the monitoring scene requirements: high privacy scenes output LoP1, medium privacy scenes output LoP2 or LoP3, and real-time and storage-constrained scenes prioritize LoP1 or LoP2.

2. The intelligent video surveillance management method based on big data as described in claim 1, characterized in that: The process of acquiring and preprocessing the raw video stream to generate standardized low-resolution images and human bounding boxes involves using Gaussian filtering on the raw frames. Denoising is performed to obtain smooth frames. The current frame is calculated using the inter-frame difference method. With the previous frame If the sum of the pixel differences is less than a threshold k, it is marked as a static frame, only its timestamp is recorded, and subsequent processing is skipped. If the sum of the differences is greater than or equal to the threshold k, it is marked as a non-static frame. Scale to 512×512 resolution to generate a standardized image. Using the YOLOv5s model in normalized images Detects human bodies and outputs the coordinates of the human body bounding box. and confidence level Calculate the normalized image of the current frame. With the previous frame The pixel differences are used to generate a difference image. Then, a threshold A is applied to binarize the difference image to generate a dynamic region mask. Regions with pixel differences greater than or equal to threshold A are marked as 1, representing moving regions, while regions with pixel differences less than threshold A are marked as 0, representing static regions.

3. The intelligent video surveillance management method based on big data as described in claim 2, characterized in that: Based on standardized low-resolution images and human bounding boxes, 2D keypoint coordinates and their confidence scores of the human body and face are extracted using morphological thinning algorithms and the MTCNN model. The average confidence score of facial keypoints is calculated, and the facial keypoints and confidence scores are dynamically updated. The standardized images are then processed using morphological thinning algorithms. middle For the corresponding region, extract the human skeleton image to represent the human centerline structure. Use Hough transform to detect straight line segments in the skeleton image to identify limb parts. Locate 25 key points through geometric constraints, calculate the confidence score for each key point, and generate the key point location. and confidence vector ,according to Trim the human body ROI region ; After cutting 68 facial landmarks were extracted using the MTCNN model to generate facial landmarks. and confidence vector Calculate the mean confidence score of facial key points. If mean If the threshold Q is greater than or equal to the current frame's threshold, then the current frame's threshold is used directly. As output, input to the DECA model, if mean <Threshold Q, extract dynamic region mask For regions with a median value of 1, rerun MTCNN within the masked dynamic region to obtain new facial key points. and confidence level , use new As input to the DECA model, if the new mean If the result is greater than or equal to the threshold Q, use the new result; otherwise, use the facial landmarks and confidence scores from the previous frame.

4. The intelligent video surveillance management method based on big data as described in claim 3, characterized in that: Based on human keypoints and confidence levels, the VIBE model is used to generate SMPL-X model parameters, obtaining the 3D human head vertex coordinates. Based on dynamically updated facial keypoints and confidence levels, the DECA model is used to generate FLAME parameters, obtaining the 3D face vertex coordinates. The vertex coordinates of the human and face meshes are then fused, and the SMPL-X and FLAME parameters are compressed. The VIBE model is then used based on the human keypoint positions. and confidence vector Generate keypoint rotation parameters for the SMPL-X model Shape parameters Translation parameters Rotate keypoint parameters Shape parameters Translation parameters Input an SMPL-X model, calculate the vertex positions of the 3D mesh based on the forward kinematics function, and output the vertices of the 3D human body mesh. For rotation parameters Perform time series optimization and calculate smoothing rotation parameters. Extracting fixed points of the head region from the vertices of a 3D human body mesh Coordinates, based on the predefined head vertex index i of the SMPL-X model, generate a fixed point coordinate set, denoted as ; Using the DECA model, based on 68 dynamically updated facial landmarks using MTCNN. and confidence level Generate FLAME model parameters, including facial pose. facial shape Expression parameters FLAME model outputs 3D facial mesh vertices Update facial shape parameters Based on 68 facial key points output by the MTCNN model, and utilizing the predefined topology of the FLAME model, a subset of 3D mesh vertices corresponding to the 68 key points is obtained. Through the key point-to-vertex index mapping provided by FLAME, the core vertices of the key facial regions are selected. The weighted average of the 3D coordinates of the core vertices is calculated to generate the facial vertex coordinates. ; By using the predefined SMPL-X head vertex index and FLAME model face vertex mapping, a system is established. The correspondence between the head vertices and the FLAME mesh vertices is determined using the iterative nearest-point algorithm. The rigid transformation matrix of the FLAME mesh vertices is calculated to ensure... facial apex and The head vertices are aligned, and the aligned FLAME mesh vertices and SMPL-X head mesh vertices are weighted and fused to generate new 3D mesh vertices. The merged head apex The body mesh vertices are merged with those of SMPL-X, and the Laplacian smoothing algorithm is applied to optimize the mesh geometric continuity, generating a unified 3D human body model. Differential coding is applied to the SMPL-X and FLAME parameters to record inter-frame parameter changes, and dynamic region masks are used. Flattened into a 1D binary sequence, the input is the ZStandard compression algorithm. The binary sequence is compressed using dictionary compression and entropy coding mechanisms to generate compressed mask data. Huffman coding is then used to perform secondary compression on the differencing parameters and the ZStandard compressed mask to generate smaller anonymized data packets.

5. The intelligent video surveillance management method based on big data as described in claim 4, characterized in that: The generation of semantic data compression and encryption data packets based on semantic feature vectors refers to using MeshLab to process new 3D mesh vertices. Vertex simplification is performed to generate compressed mesh data. The object mask is flattened into a 1D bit string, and a compressed object mask is generated using the Brotli compression algorithm. The object's 3D positioning coordinates, local 3D scene features, and semantic feature vectors are quantized into 16-bit floating-point numbers to form a semantic parameter vector. The semantic parameter vector, compressed object mask, and object category label are packaged into Protobuf format data. The Protobuf data is encrypted using the XChaCha20 encryption algorithm to generate an encrypted data packet. The encrypted data packet is transmitted to the cloud server via a 5G network using the gRPC protocol.

6. A big data-based intelligent video surveillance management system, based on the big data-based intelligent video surveillance management method according to any one of claims 1 to 5, characterized in that: This includes a video preprocessing and human detection module, which is used to acquire raw video streams and preprocess them to generate standardized low-resolution images and human bounding boxes; The key point extraction and dynamic update module is used to extract 2D key points and confidence scores of the human body and face based on standardized images and bounding boxes. The 3D modeling and parameter compression module is used to generate SMPL-X and FLAME parameters based on 2D key points, output 3D human body and face mesh vertices, and compress parameters; The object localization and semantic analysis module is used to generate 3D localization and semantic features of objects based on dynamic region masks and 3D mesh vertices; The data compression and encrypted transmission module is used to compress 3D mesh vertices and semantic features and encrypt them for transmission to the cloud.

7. 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 video surveillance management method based on big data as described in any one of claims 1 to 5.

8. 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 video surveillance management method based on big data as described in any one of claims 1 to 5.

Images