Multi-modal dual-person pose modeling and privacy stick figure generation method and system

Abstract

The application provides a multi-modal two-person posture modeling and privacy stick figure generation method and system. The acquisition end closes the RGB and collects the structured light depth map, radar point cloud and micro-Doppler data; the two-way data level fusion is performed to fill the depth holes with the radar and filter the radar clutter with the human body mask; the gating micro-Doppler identity correlation cost is constructed, and the two persons are matched with the minimum total cost; the two-mode feature input is introduced into the customized fusion model of the skeletal topology position coding, the micro-Doppler additional channel, the confidence mask and the identity gating, and the refined three-dimensional skeleton is output, and the occluded joints are completed in combination with the radar speed and the kinematics constraint; then, the privacy level is determined according to the sensitive area, the sensitive action and the two-person contact through the hysteresis switching, and the two-person stick figure or the event level result is output in stages. The application realizes the identity stability, the accurate occlusion completion and the high-precision three-dimensional modeling in the two-person close scene, and takes into account the supervision continuity and the privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare and indoor multimodal perception technology, specifically to a method and system for multimodal dual-person pose modeling and privacy stick figure generation. Background Technology

[0002] In indoor two-person interaction scenarios such as long-term care insurance services, elderly care, and home rehabilitation, effective monitoring of caregiving activities is necessary while protecting the privacy of those being cared for. Traditional video camera monitoring methods collect a large amount of private information, including real faces, physical appearance, clothing, and home environment. This is particularly unacceptable during sensitive processes such as assisting with dressing, washing and care, toileting, bedpan cleaning, and turning over.

[0003] Existing single-vision or infrared structured light solutions are prone to depth holes, instance segmentation errors, and joint misidentification in scenarios with strong light, smoke, reflective materials, limb occlusion, or close proximity of two people. While single 4D millimeter-wave radar solutions are privacy-friendly and can measure speed, their point clouds are sparse, making it difficult to accurately model fine-grained joints such as wrists, elbows, knees, and ankles, and also difficult to generate high-quality stick figure images.

[0004] Existing radar and depth camera fusion solutions are mostly one-way fusion or feature-level / decision-level fusion, lacking an effective data complementarity mechanism in close-range two-person interaction scenarios. Although Micro-Doppler features have been used for person recognition and gait analysis, their ability to stably track identities is limited in scenarios where two people have crossed limbs, are in the initial stage of initiation, or are both people simultaneously stationary. Existing joint occlusion processing methods mostly use simple interpolation or historical trajectory extrapolation, which are difficult to combine with radar dynamic information and human kinematic constraints.

[0005] Furthermore, existing methods of anonymizing skeletons or stick figures typically employ fixed rules (such as always hiding the head or always showing the complete torso), failing to dynamically adjust the desensitization intensity based on sensitive areas, sensitive actions, and the relationship between two individuals. Fixed privacy modes either fail to meet regulatory continuity requirements or cannot effectively prevent the inference of body features and fine motor trajectories. Therefore, there is an urgent need for a multimodal technical solution that can achieve stable tracking of two individuals' identities, intelligent completion of occluded joints, and strike a balance between regulatory explainability and dynamic privacy protection. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes a multimodal dual-person pose modeling and privacy stick figure generation method and system. Through bidirectional data-level complementary fusion, gated Micro-Doppler identity association, and customized multimodal feature fusion, it achieves high-precision 3D skeleton modeling and occlusion completion in complex indoor dual-person interaction scenarios. Through a three-level dynamic privacy mechanism triggered by sensitive area, sensitive action recognition, and dual-person contact relationship, as well as a specific torso desensitization method, it effectively protects privacy while preserving the continuity of coarse-grained supervision.

[0007] To achieve the above objectives, the present invention adopts the following technical solution, including: S1 acquires infrared structured light depth map, infrared map, radar point cloud and micro-Doppler data under minimized control at the acquisition end, and shuts down or isolates RGB image stream during normal operation; S2 performs filtering, back projection, and human instance segmentation on the structured light data to extract key points of the 3D skeleton of two people, joint confidence, and human mask. S3 performs filtering and static background removal on radar point clouds, performs dual-person clustering and trajectory tracking based on spatial position and velocity, and constructs gated regions for each target to extract micro-Doppler time-frequency features. S4 performs time synchronization and spatial registration, mapping the structured light skeleton, mask, radar point cloud, and trajectory to a unified world coordinate system; S5 performs bidirectional data-level fusion, using radar points to fill in structured light depth holes and using structured light human body masks to remove radar clutter points; S6. Construct an identity association cost function that includes centroid, velocity, size, historical trajectory and micro-Doppler differences, and determine the correspondence between structured light instances and radar trajectories by matching the minimum total cost. S7, for each target after identity association, introduces the structured light side and radar side feature inputs into a customized multimodal feature fusion model of skeleton topological position encoding, micro-Doppler additional channel, confidence mask and identity gating, and outputs a refined three-dimensional skeleton; S8 determines occluded or missing joints based on confidence level, depth validity and historical trajectory, and completes them by combining radar velocity direction, micro-Doppler trend and kinematic constraints and adds completion labels. S9 determines the current privacy level based on sensitive areas, sensitive actions, duration of contact between two people, and external commands, using dynamic switching logic with hysteresis mechanism; S10, output according to the current privacy level: Level 1 outputs anonymized two stick figures, Level 2 hides the head, desensitizes the torso, quantifies the ends and reduces the frequency, Level 3 only outputs event-level results.

[0008] Furthermore, the customized multimodal feature fusion model described in S7 is constructed and run as follows: the skeleton of each human body is represented as a graph structure G=(V,E) consisting of joint nodes and skeletons, and a topological position code PEtopo is generated for each joint node based on joint type, left and right side identifiers and topological distance on the graph, in order to distinguish joints with different semantics but similar spatial appearance. On the structured light side, the joint features, the topological position encoding, and the joint confidence are concatenated to form a query vector Q; On the radar side, the micro-Doppler time-frequency characteristics within the gated area are incorporated into the key vector K and value vector V as an additional micro-Doppler channel to enhance the perception of human motion rhythm. Apply a joint confidence mask to joints with confidence scores below a threshold to increase the attention weight of the joint to the radar compensation branch, so that radar features can compensate for the absence or unreliability of structured light joints. An identity gating mask is applied to the cross-attention weights so that each structured light joint only focuses on the key value corresponding to the radar gating region that matches its identity, thereby suppressing the cross-contamination of radar features between two people. Residual connections are introduced during the fusion process to preserve the high-precision priors of the structured light joints and stabilize training. The system employs multi-head cross-attention to calculate and fuse attention across four subspaces: spatial location, velocity direction, micro-Doppler time-frequency texture, and historical trajectory. Finally, the regression head outputs the refined 3D coordinates and fusion confidence of each joint, enabling fine-grained joints such as the wrist, elbow, knee, and ankle to achieve higher localization accuracy than general attention models even under sparse radar point cloud conditions. This provides a reliable basis for subsequent occluded joint identification and completion.

[0009] Furthermore, the bidirectional data-level fusion described in S5 also includes: determining the structured light modal uncertainty based on the structured light depth map hole rate and the integrity of the human body mask. Radar mode uncertainty is determined based on the radar point cloud density and the mean signal-to-noise ratio. ,according to and The fusion weights are dynamically allocated and weighted according to the weights in deep hole completion and radar clutter filtering.

[0010] Furthermore, the micro-Doppler differences described in S6 are obtained by extracting lengths of [value missing] from each human body within the gated region. The sliding time window micro-Doppler time-frequency sequence is obtained by calculating its time-frequency texture similarity. When the length of historical data is less than the minimum effective window If the micro-Doppler energy is below the threshold, the weight of the micro-Doppler term is dynamically reduced or set to zero, degenerating into an association strategy that mainly relies on the centroid, size, and historical trajectory.

[0011] Furthermore, the completion in S8 includes: when a joint meets the conditions of confidence level below a threshold, invalid corresponding depth, or missing multiple consecutive frames, the velocity direction of the historical trajectory extrapolation position is corrected by combining the radar local velocity direction vector and micro-Doppler motion trend, while applying human skeleton length constraints and left and right limb symmetry constraints to generate the completed joint coordinates and completion confidence.

[0012] Furthermore, the sensitive actions described in S9 are identified based on a continuous three-dimensional skeletal sequence within a time window, the distance between the centroids of the two individuals, the limb contact relationship, and the area where the human body is located; When the confidence level of a sensitive action is not lower than the first threshold θs and the upgrade frame count is met consecutively, the privacy level is upgraded from Level 1 to Level 2; when the confidence level of a highly sensitive action is not lower than the second threshold θh and the upgrade frame count is met consecutively, the level is upgraded to Level 3.

[0013] Furthermore, the Level 2 trunk desensitization includes at least one of the following methods: Method A: Calculate the midpoint of the shoulder and the midpoint of the hip, render only the center line segment between them, and hide the chest width, waist width and body curves; Method B replaces the quadrilateral formed by the left and right shoulders and the left and right hips with a standardized rectangle or trapezoid, and performs interval quantization on the shoulder width and hip width in steps of 5–15cm. Method C performs downsampling and curve simplification on the torso contour point set, retaining only the approximate shape of the area occupied by the human body; Method D divides the torso region into low-resolution grids, with each grid retaining only the average depth or occupancy state, and rendering it as a blocky or blurred area.

[0014] Furthermore, the hysteresis mechanism is set as follows: upgrading from Level 1 to Level 2 requires 10 consecutive frames, downgrading from Level 2 to Level 1 requires 30 consecutive frames with a minimum hold time of 2 seconds, upgrading from Level 2 to Level 3 requires 10 consecutive frames, and downgrading from Level 3 to Level 2 requires 60 consecutive frames with a minimum hold time of 5 seconds; the 10, 30, and 60 frames are all fixed frame number thresholds, and under variable frame rate conditions, the fixed frame number thresholds are automatically converted into corresponding time thresholds.

[0015] A multimodal dual-person pose modeling and privacy stick figure generation system is provided to implement the aforementioned multimodal dual-person pose modeling and privacy stick figure generation method. The system includes a data acquisition module, a data minimization module at the acquisition end, a structured light preprocessing module, a radar preprocessing module, a spatiotemporal alignment module, a bidirectional data-level fusion module, a dual-person identity association module, a customized multimodal feature fusion module, an occlusion joint completion module, a sensitive region and action recognition module, a dynamic privacy level determination module with hysteresis, a hierarchical desensitization rendering module, and a stick figure and event-level output module.

[0016] Furthermore, the system also includes an infrared structured light device and a 4D millimeter-wave radar. The infrared structured light device is used to collect depth data of indoor scenes, and the 4D millimeter-wave radar is used to collect human point clouds, radial velocity, signal-to-noise ratio, and micro-Doppler slow time series. The data acquisition module is used to acquire infrared structured light data and 4D millimeter-wave radar data; The data minimization module at the acquisition end is used to shut down, filter, or isolate unnecessary RGB image streams; The structured light preprocessing module is used to filter the depth map, segment human instances, extract skeletal key points, and extract human contours. The radar preprocessing module is used to filter, perform two-person clustering, target tracking, and extract micro-Doppler features from the radar point cloud. The spatiotemporal alignment module is used to synchronize the time and unify the spatial coordinates of the two types of data. The bidirectional data-level fusion module is used to complete radar point cloud filling of structured light depth holes and structured light masking to constrain radar clutter. The dual-person identity association module is used to match dual-person identities based on location, speed, size, micro-Doppler features, and historical trajectory. The customized multimodal feature fusion module is used to output a refined 3D skeleton through a multimodal fusion model; The occluded joint completion module is used to intelligently complete low-confidence, missing, or occluded joints. The hierarchical desensitization rendering module is used to perform head hiding, torso simplification, contour quantization, raster blurring and end joint quantization according to the privacy level. The sensitive area and action recognition module includes a sensitive area recognition module and a sensitive action recognition module; The sensitive area identification module is used to determine whether a human body has entered a preset sensitive area; The sensitive action recognition module is used to identify sensitive actions based on continuous skeletal sequences, distance between two people, and limb contact relationships. The dynamic privacy level determination module with hysteresis includes a privacy level determination module and a privacy level switching module; The privacy level determination module is used to calculate the current privacy level based on the basic level, sensitive areas, sensitive actions, contact relationships, and external instructions; The privacy level switching module is used to perform privacy level upgrades and downgrades according to the hysteresis mechanism; The stick figure and event-level output module includes a stick figure generation module and a data output module. The stick figure generation module is used to generate a two-person stick figure image, video stream, or event-level result based on the skeletal topology. The data output module is used to output the de-identified result corresponding to the current privacy level.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention provides a method and system for multimodal dual-person pose modeling and privacy stick figure generation. Through bidirectional data-level complementary fusion (radar point cloud completion of structured light depth holes + structured light human body mask filtering radar clutter), it provides clean and reliable input data for subsequent identity association and feature fusion, significantly improving the processing stability in dual-person close-up and limb-crossing scenarios.

[0018] 2. This invention provides a multimodal dual-person pose modeling and privacy-preserving stick figure generation method and system. It combines gated Micro-Doppler temporal features with a degradation strategy to solve the identity transition problem in the initial startup phase and in static dual-person scenes in existing technologies. Simultaneously, Micro-Doppler is also used for motion trend prediction of occluded joints, forming a collaborative feature reuse mechanism. A customized cross-attention model distinguishes semantically different but visually similar joints through topological position encoding, prevents cross-contamination of dual-person radar features through identity gating, and enhances motion rhythm perception through additional channels of Micro-Doppler, achieving refined skeletal output superior to general fusion models.

[0019] 3. This invention provides a method and system for multimodal dual-person pose modeling and privacy stick figure generation, and a radar speed and Micro-Doppler-assisted kinematic constraint completion method, which makes the occluded joint recovery results more consistent with the actual human movement law.

[0020] 4. This invention provides a method and system for multimodal dual-person pose modeling and privacy stick figure generation. The three-level dynamic privacy mechanism combines sensitive area triggering, sensitive action recognition (based on pose sequence + contact relationship) and dual-person distance. Through specific Level 2 torso geometry desensitization methods (centerline replacement, quadrilateral regularization, contour downsampling, grid blurring, etc.) and explicit hysteresis parameters, it effectively reduces the risk of inferring body shape features and fine action trajectories while preserving coarse-grained action coherence and regulatory interpretability. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is the overall system architecture diagram of the present invention; Figure 2 This is the overall flowchart of the present invention; Figure 3 This is a privacy control relationship diagram of the acquisition end, processing intermediate state, and output end of this invention; Figure 4 This is a sequence diagram of the dual-person identity association in this invention; Figure 5 This is a schematic diagram of the multimodal feature fusion model of the present invention; Figure 6 This is a dynamic privacy level switching state diagram of the present invention; Figure 7 This is a schematic diagram of the stick figure generation interface of the present invention; In the diagram: 110, Infrared structured light device; 120, 4D millimeter-wave radar; 210, Data acquisition module; 220, Data minimization module at the acquisition end; 310, Structured light preprocessing module; 320, Radar preprocessing module; 410, Spatiotemporal alignment module; 420, Bidirectional data-level fusion module; 430, Dual-person identity association module; 440, Customized multimodal feature fusion module; 450, Occlusion joint intelligent completion module; 510, Sensitive area recognition module; 520, Sensitive action recognition module; 610, Privacy level determination module; 620, Privacy level switching module; 710, Hierarchical desensitization rendering module; 720, Stick figure generation module; 730, Data output module. Detailed Implementation

[0023] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0024] See Figure 1This embodiment provides a multimodal indoor dual-person posture modeling and dynamic privacy stick figure generation system based on 4D millimeter-wave radar and infrared structured light. The system includes two types of front-end acquisition devices: an infrared structured light device 110 and a 4D millimeter-wave radar 120, as well as multiple processing modules running in a controlled computing environment. The infrared structured light device 110 is used to acquire depth maps, infrared maps, and optional color images of the indoor scene, while the 4D millimeter-wave radar 120 is used to acquire human point clouds, radial velocity, signal-to-noise ratio, range-velocity distribution, and micro-Doppler slow time series.

[0025] Along the data flow direction, the system sequentially includes: a data acquisition module 210 for acquiring infrared structured light data and 4D millimeter-wave radar data; a data minimization module 220 for shutting down, filtering, or isolating unnecessary RGB image streams during the acquisition phase; a structured light preprocessing module 310 for filtering depth maps, segmenting human instances, extracting skeletal key points, and extracting human contours; a radar preprocessing module 320 for filtering radar point clouds, performing two-person clustering, target tracking, and micro-Doppler feature extraction; a spatiotemporal alignment module 410 for achieving time synchronization and spatial coordinate unification of the two types of data; a bidirectional data-level fusion module 420 for completing radar point cloud filling of structured light depth holes and structured light masking to constrain radar clutter; a two-person identity association module 430 for performing two-person identity matching based on position, velocity, size, micro-Doppler features, and historical trajectory; and a customized multimodal feature fusion module 440 for performing multimodal feature fusion... The system includes: a model outputting a refined 3D skeleton; an intelligent joint completion module 450 for intelligently completing low-confidence, missing, or occluded joints; a sensitive area recognition module 510 for determining whether a human body has entered a preset sensitive area; a sensitive action recognition module 520 for recognizing sensitive actions based on continuous skeleton sequences, distance between two people, and limb contact relationships; a privacy level determination module 610 for calculating the current privacy level based on the basic level, sensitive areas, sensitive actions, contact relationships, and external commands; a privacy level switching module 620 for performing privacy level upgrades and downgrades based on a hysteresis mechanism; a graded desensitization rendering module 710 for performing head hiding, torso simplification, contour quantization, raster blurring, and end joint quantization based on the privacy level; a stick figure generation module 720 for generating two-person stick figure images, video streams, or event-level results based on the skeleton topology; and a data output module 730 for outputting desensitization results corresponding to the current privacy level.

[0026] See Figure 2The overall process of this embodiment includes steps S1 to S10. The preferred embodiments of each step are described in detail below with reference to each module. This invention divides privacy protection into three levels: minimum control at the acquisition end, isolation control in the intermediate processing state, and dynamic hierarchical desensitization control at the output end, and these levels are integrated throughout the following steps.

[0027] like Figure 3 As shown, the present invention establishes a hierarchical privacy control relationship between the acquisition end, the intermediate processing state, and the output end. The acquisition end performs minimal control on the RGB image stream by closing, filtering, or isolating it. The intermediate processing state performs short-term isolation processing on the depth map, infrared map, radar point cloud, and micro-Doppler features in a controlled computing environment and does not output the original data to the outside by default. The output end performs dynamic hierarchical desensitization on the stick figure image and event-level results according to the current privacy level.

[0028] S1. Acquire minimized infrared structured light data and 4D millimeter-wave radar data. The data acquisition module 210 acquires depth data of the indoor scene through the infrared structured light device 110. Its input data can be represented as: Frame RGBD = { depth: H×W depth map; ir: H×W infrared image, optional; RGB: H×W×3 image, optional; K: Camera intrinsic parameter matrix; T cam Camera extrinsic parameter matrix; timestamp: Collection timestamp } The camera intrinsic parameter matrix can be represented as:

[0029] For a pixel coordinate (u, v) in the depth map, with a corresponding depth value of Z, the 3D coordinates of that pixel in the camera coordinate system are: ; ;

[0030] The data acquisition module 210 also acquires human point clouds, radial velocity, signal-to-noise ratio, range-velocity distribution, and micro-Doppler slow time series data via the 4D millimeter-wave radar 120. The radar point data can be represented as: Point4D = { x: Three-dimensional coordinate x; y: Three-dimensional coordinate y; z: Three-dimensional coordinate z; v_r: Radial velocity; SNR: Signal-to-noise ratio; timestamp: timestamp } In normal operation, the data minimization module 220 preferably shuts down the RGB image stream at the acquisition end, or prevents RGB images from entering the human posture modeling module and the external output module. The human posture modeling of this invention preferably uses depth maps, infrared maps, radar point clouds, radar velocity, micro-Doppler features, timestamps, and calibration parameters; RGB images are not a necessary input for completing human posture modeling.

[0031] If the infrared structured light device 110 has a built-in RGB channel, the data minimization module 220 at the acquisition end can control it in at least one of the following ways: turn off the RGB image stream at the device driver layer; do not create an RGB image cache; do not pass the RGB image to the subsequent attitude modeling module; do not transmit the RGB image to the external system; do not write the RGB image to persistent storage; only temporarily enable the RGB image stream in authorized debugging or calibration state, and turn it off immediately after completion.

[0032] Depth maps, infrared maps, radar point clouds, and micro-Doppler features are intermediate data required for pose modeling. To ensure the effectiveness of occlusion completion, identity association, and sensitive action recognition, this invention allows for short-term processing of the above data in a local or controlled computing environment. Preferably, intermediate data is stored in volatile storage space, retaining only the sliding window data required for real-time calculation. Data exceeding the preset time window is automatically released. The original depth maps, infrared maps, and radar point clouds are not output to external systems by default, nor are they written to long-term storage by default.

[0033] S2. Preprocess the infrared structured light data, segment human instances, and extract skeletal key points. The processing flow of the structured light preprocessing module 310 includes depth map filtering, invalid depth removal, human region recognition, human instance segmentation, skeletal key point detection, 3D coordinate back projection, human contour extraction, joint visibility calculation, and joint confidence calculation.

[0034] The structured light preprocessing module 310 performs bilateral filtering on the depth map: ; Where D(p) represents the original depth value of pixel p. Represents the neighborhood of pixel p. Indicates spatial distance weights. Indicates depth difference weights. Represents the normalization coefficient. This represents the filtered depth value. Furthermore, temporal median filtering can be performed on multiple consecutive depth maps to reduce speckle jitter.

[0035] The structured light preprocessing module 310 backprojects the depth map into a 3D point cloud and performs instance segmentation of the human body based on spatial distance, human height, human volume, and connectivity. In a two-person scene, the system retains the two main human body instances: ; in, This represents the mask for the first human instance. This represents the mask for the second human instance. For each human instance region, extract the key points of the human skeleton. For the m-th human target, its set of key points can be represented as: ; in, In the formula , , This represents the three-dimensional coordinates of the k-th joint. This represents the confidence level of the k-th joint, where K represents the number of joints.

[0036] This invention may utilize, but is not limited to, the key points of the human skeleton shown in Table 1: Table 1

[0037] The connections between the human skeleton can include:

[0038] The above-described skeletal topology is merely an example, and the present invention does not limit the specific number of joints or the connection method.

[0039] S3. Perform point cloud filtering, two-person clustering, target tracking, and micro-Doppler feature extraction on the 4D millimeter-wave radar data. The radar preprocessing module 320 performs height range filtering, velocity range filtering, signal-to-noise ratio filtering, static background filtering, outlier removal, and preliminary multipath reflection filtering on the radar point cloud. The filtered radar point cloud can be represented as: ; in, This represents the point cloud filtering function.

[0040] The radar preprocessing module 320 performs human body clustering on the radar point cloud based on the spatial position, velocity, and signal-to-noise ratio distribution. The clustered radar human body clusters can be represented as follows: ; in, This represents the first human body radar point cloud cluster. Let represent the second human body radar point cloud cluster. For each human body cluster, calculate the cluster centroid, height, volume, velocity vector, point cloud density, and signal-to-noise ratio distribution. For the nth radar human target, its state vector can be represented as:

[0041] Prediction is performed using a state transition model:

[0042] The observation-updated model is used for correction: ; Where F represents the state transition matrix, Indicates the filter gain. H represents the current observation value, and H represents the observation matrix. The radar preprocessing module 320 also maintains a slow-time radar sequence for each human target and extracts micro-Doppler features:

[0043] Further extraction of motion frequency distribution, limb swing frequency, gait frequency, respiratory rhythm, energy distribution, and time-frequency texture features.

[0044] S4. Perform time synchronization and spatial registration on the two types of data. The time-space alignment module 410 completes the time synchronization and spatial registration of the two types of data. Let the timestamp of the structured light data frame be... The radar data frame timestamp is The maximum allowable time difference is .

[0045] When the following conditions are met:

[0046] It is then assumed that the two frames of data can be directly matched; When the following conditions are met:

[0047] Then, interpolation or motion extrapolation is performed based on the radar trajectory status: ; in, Indicates the reference time. For spatial registration, the world coordinate system is defined as W, the infrared structured light coordinate system as C, and the radar coordinate system as R. Structured light point transformation to the world coordinate system:

[0048] Radar point conversion to world coordinate system:

[0049] The extrinsic transformation matrix is: ; In the formula, R represents the rotation matrix and t represents the translation vector. Through spatial registration, the spatiotemporal alignment module 410 maps the structured light skeleton points, human contour points, radar point clouds, and radar trajectories to the world coordinate system.

[0050] S5. Perform bidirectional data-level fusion. The bidirectional data-level fusion module 420 performs data-level fusion including radar point cloud completion of structured light depth holes, structured light human body mask constraint of radar point clouds, and dynamic adjustment of data fusion weights based on modal confidence. When invalid depth pixels appear in the human body region of the structured light depth map, the radar point cloud is projected onto the structured light image space or world coordinate space; for pixels (u,v) in the depth hole region, the nearest radar point is searched: ; in, This represents the spatial ray corresponding to the pixel. This represents a radar point. When the nearest distance is less than a preset threshold, the radar point depth information is used to fill in the gaps in the area.

[0051] The bidirectional data-level fusion module 420 also projects radar points onto the structured light human body mask space. For any radar point, if the point is not located within the human body mask or the human body mask expansion area, it is marked as a candidate clutter point. If the point further meets at least one of the following conditions: velocity below a threshold, signal-to-noise ratio below a threshold, remaining stationary for a long time, distance from the human body centroid too large, inconsistent with the human body contour, or inconsistent with the historical trajectory, then the radar point is removed.

[0052] Regarding adaptive mode weights, let the structured light mode uncertainty be... The radar mode uncertainty is The weights of structured light can be expressed as:

[0053] Radar weights can be expressed as:

[0054] When the quality of the structured light depth map deteriorates, the radar mode weight is increased; when the radar point cloud is sparse or the signal-to-noise ratio decreases, the structured light mode weight is increased, thereby achieving bidirectional complementarity between the two modes.

[0055] S6. Perform dual-person identity association. The dual-person identity association module 430 matches the human instance identified by structured light with the human target tracked by radar in each frame. For the structured light human instance i and the radar target j, the association cost is constructed as follows: ; in, This indicates the difference in the position of the human body's center of mass. This indicates the difference between the direction and magnitude of velocity. This indicates differences in human height, shoulder width, and trunk dimensions. Indicates differences in micro-Doppler characteristics. Indicating differences in historical trajectories, This indicates the weight of each item.

[0056] Regarding micro-Doppler similarity, let the micro-Doppler feature vectors of the two human targets be respectively... and The cosine similarity is:

[0057] The cost of micro-Doppler is:

[0058] Based on the identity association cost matrix:

[0059] The dual-person identity association module 430 uses a minimum total cost matching algorithm to determine the identity correspondence:

[0060] This method can reduce the probability of identity switching in scenarios involving two people crossing, being close, obscuring each other, supporting each other, and having overlapping limbs.

[0061] like Figure 4 As shown, Figure 4 The present invention illustrates the dual-person identity association by illustrating the crossover process of the centroid trajectories of two individuals, ID1 and ID2, over time: the centroid trajectories of the two individuals, ID1 and ID2, undergo three stages: before the crossover, during the crossover (short-term aliasing), and after the crossover. Within the short-term aliasing region caused by the crossover, the present invention constructs a gated region based on distance, angle, and velocity, and combines it with a micro-Doppler sliding time window of length T_MD, applying an identity association cost weighted by five factors: position, velocity, size, micro-Doppler, and historical trajectory. By performing minimum total cost matching, stable identity correspondence can be maintained and identity jumps can be reduced in situations where two people cross paths and are in close proximity.

[0062] S7. Perform multimodal feature-level fusion to generate a refined 3D human skeleton. The customized multimodal feature fusion module 440 encodes both structured light-side features and radar-side features. Structured light-side features include 3D joint coordinates, joint confidence, joint visibility, depth local blocks, human contour, human mask, body scale, human centroid, and human orientation. The structured light feature encoding is represented as follows: ; Where J represents skeletal keypoints, D represents depth information, M represents the human body mask, and C represents joint confidence. This represents a structured light encoder. Radar-side features include point cloud spatial distribution, radar human clusters, centroid position, velocity vector, local velocity field, signal-to-noise ratio distribution, micro-Doppler features, and historical trajectory. The radar feature encoding is represented as follows: ; Where P represents radar point cloud, V represents velocity feature, MD represents micro-Doppler feature, and H represents historical trajectory. This refers to a radar encoder.

[0063] Furthermore, the overall structure of the customized multimodal feature fusion model of this invention is as follows: Figure 5 As shown, Figure 5 The diagram illustrates a structured light side-encoding branch (consisting of a query vector Q made up of joint features, topological location encoding PEtopo, and joint confidence), a radar side-encoding branch (consisting of a key K and a value V made up of micro-Doppler time-frequency features within the gated region as additional micro-Doppler channels), a multi-head cross-attention fusion structure, and a regression head that outputs refined 3D skeletal coordinates and fusion confidence. The customized multimodal feature fusion model is constructed and run as follows: Each human skeleton is represented as a graph structure consisting of joint nodes and bone connections.

[0064] A topological location code is generated for each joint node based on joint type, left and right side identifiers, and topological distance on the graph to distinguish joints with different semantics but similar spatial appearance.

[0065] In cross-attention calculation, Generate queries using structured light joint features , Key generation based on radar features Sum And adaptively generate modal fusion weights for each joint. This allows for joint-by-joint customized weighted fusion of structured light and radar modes:

[0066] A multi-head attention mechanism is used to model the modal relationships between different subspaces in parallel:

[0067] The fusion results from each head are concatenated and mapped to obtain the fused features:

[0068] In the overlapping area of ​​two people, a mask is introduced. Attention weights Constraints are applied to suppress erroneous associations with other human joints. Structured light features and radar features are fused using a cross-attention mechanism after being mapped to a unified feature space. ; Where d represents the feature dimension. The customized multimodal feature fusion module 440 outputs a refined skeleton through the regression head:

[0069] Refined skeleton includes: ; Where xk, yk, and zk represent the three-dimensional coordinates of the k-th joint, and ck represents the fusion confidence of the k-th joint.

[0070] S8. Intelligent completion of occluded joints. The intelligent occlusion joint completion module 450 determines a joint as occluded when it meets at least one of the following conditions: confidence level below a threshold, invalid corresponding depth, missing data in multiple consecutive frames, significant difference between coordinates and historical trajectory, violation of human kinematic constraints, or location within an overlapping area of ​​two human bodies. Occlusion joint completion incorporates at least one or more of the following: historical skeleton sequence, radar velocity trend, left-right limb symmetry, geometric relationship between adjacent joints, human skeleton length constraints, relative position of two people, human action type, and prediction results from a human posture temporal model.

[0071] Taking the shoulder joint (S), elbow joint (E), and wrist joint (W) as examples, when the elbow joint is obscured, the position of the elbow joint is estimated based on human kinematic constraints: ; in, This represents the direction vector from the shoulder joint to the wrist joint. This represents the normal vector of the limb's motion plane, where α and β are parameters determined based on the length of the upper arm, forearm, and historical posture. If radar velocity characteristics indicate a clear direction of movement for the limb, further adjustments are made based on the radar velocity. ; in, This represents the local velocity vector estimated by the radar. This represents the velocity correction factor. For continuously occluded joints, a temporal prediction model can be used: ; Among them, historical skeletal sequences, velocity sequences, and micro-Doppler sequences are used as inputs to the time-series prediction model G. This indicates the occluded joint after completion.

[0072] S9. Generate a two-person stick figure image, video stream, or event-level result based on the dynamic privacy level. The privacy level determination module 610 pre-sets a basic privacy level. The basic privacy level can be configured based on application scenarios, room types, user authorization status, service types, regulatory requirements, time periods, and external control policies. The privacy level determination module 610 determines whether to upgrade the privacy level based on at least one of the following conditions: sensitive area triggering, sensitive action triggering, close contact between two people triggering, occlusion of sensitive areas of the human body triggering, user active control command triggering, regulatory system control command triggering, and authorization status change triggering.

[0073] The sensitive area identification module 510 predefines a set of sensitive areas in the world coordinate system:

[0074] Sensitive areas may include bedside care areas, bathroom areas, washing and care areas, changing areas, and user-defined areas. When the body's center of mass, hip center, torso center, or other designated key points enter a sensitive area and remain there for more than a preset time, the privacy level will be upgraded. The sensitive area trigger function can be expressed as: ; in, This represents the coordinates of the human body's center of mass or a specified key point. The sensitive action recognition module 520 identifies sensitive actions based on continuous posture sequences, distance between two people, limb contact relationships, the area where the human body is located, and the duration of the action. The set of sensitive actions can be represented as:

[0075] For the pose sequence within the time window, the sensitive action recognition model outputs: ; in, Represents a continuous sequence of skeletal poses. Indicates the distance between two people. Indicates a physical contact relationship. This represents the area where the human body is located, and 'a' represents a candidate sensitive action. The confidence level of the sensitive action is:

[0076] When the following conditions are met:

[0077] If the preset number of frames is met consecutively, the privacy level will be upgraded to Level 2. For highly sensitive action sets:

[0078] When the following conditions are met:

[0079] Upon continuously meeting a preset frame rate, the privacy level will be upgraded to Level 3. Regarding close contact triggering between two individuals, the centroids of the caregiver and the person being cared for are set as follows:

[0080] The distance between the two people is:

[0081] When the following conditions are met:

[0082] Privacy level upgrade is triggered when at least one of the following conditions is present: overlapping torso areas, limb crossing, continuous limb contact, human entry into a sensitive area, detection of a sensitive action, or contact time exceeding a threshold. Close contact between two people does not necessarily trigger Level 3 alone; it can be used as an auxiliary condition for sensitive action recognition to avoid misjudging normal assistance as highly sensitive actions.

[0083] The privacy level determination module 610 calculates the current privacy level as the highest value among the basic level, sensitive area trigger level, sensitive action trigger level, two-person contact trigger level, and external command specified level:

[0084] like Figure 6 As shown, the present invention describes the state transition relationship of dynamic privacy levels between Level 1, Level 2, and Level 3, as well as the triggering conditions for upgrades and downgrades. The privacy level switching module 620 is equipped with a hysteresis switching mechanism: upgrading from Level 1 to Level 2 requires 10 consecutive frames (approximately 0.33 seconds at a frame rate of 30fps); downgrading from Level 2 to Level 1 requires 30 consecutive frames (approximately 1 second) with a minimum holding time of at least 2 seconds; upgrading from Level 2 to Level 3 requires 10 consecutive frames; downgrading from Level 3 to Level 2 requires 60 consecutive frames (approximately 2 seconds) with a minimum holding time of at least 5 seconds. The recommended threshold for close-range contact between two people is 0.5 meters, the recommended threshold for sensitive actions is 0.7, and the recommended threshold for highly sensitive actions is 0.8. Under variable frame rate conditions, the above fixed frame number thresholds are automatically converted into corresponding time thresholds. Through the above hysteresis mechanism, frequent switching of privacy levels caused by brief human entry into sensitive areas, instantaneous fluctuations in action recognition, or brief posture crossings are avoided, thus improving system stability.

[0085] The graded desensitization rendering module 710 outputs anonymized two-person stick figure images or video streams at Level 1. Specifically, it does not output RGB images, human body textures, clothing information, original indoor environment images, or fine-grained joints of the head and face such as eyes, ears, and nose. It can simplify the head into a single anonymous node, retain the main bones such as shoulders, elbows, wrists, hips, knees, ankles, and the center of the torso, use different colors, line types, or numbers for different human bodies, use dashed lines or semi-transparent nodes for occluded joints, and add special identifiers for algorithm-completed joints.

[0086] At Level 2, the graded desensitization rendering module 710 further reduces the recognizability of the human posture output, including at least hiding head nodes, hiding joints above the neck, simplifying the torso, blurring the human outline, performing coordinate quantization on end joints, reducing the stick figure output frequency, and outputting only direction information for some joints. The torso desensitization at Level 2 can be achieved using at least one of the following methods.

[0087] Method 1: Replace the trunk centerline. Let the left shoulder joint, right shoulder joint, left hip joint, and right hip joint be Jls, Jrs, Jlh, and Jrh, respectively, and the midpoint of the shoulder be:

[0088] The midpoint of the hip is:

[0089] Replace the original torso outline with a line segment between the midpoint of the shoulder and the midpoint of the hip:

[0090] This method no longer displays chest width, waist width, body curves, torso outlines, and body shape details, but only retains basic posture information such as standing, sitting, lying down, bending over, and tilting.

[0091] Method Two: Regularization of the Torso Quadrilateral. The torso quadrilateral is constructed based on the left and right shoulder joints and the left and right hip joints:

[0092] Replace the original quadrilateral torso with standardized rectangular, trapezoidal, or capsule-shaped regions, and quantize the shoulder width and hip width into intervals: ; Where w represents the original width, g represents the quantization step size, and Q(w) represents the quantized width. Size quantization preserves the human body's pose orientation while reducing the risk of human body shape recognition errors.

[0093] Method 3: Torso contour downsampling. Let the set of torso contour points be:

[0094] Downsampling, curve simplification, or mesh quantization are performed on the contour points. For the contour point (xi, yi), the quantization function is: ;

[0095] The quantized contour points are:

[0096] The quantized outline is only used to represent the area occupied by the human body, without preserving the fine human body boundaries.

[0097] Method 4: Raster blurring of the torso region. The torso region is divided into several grids. For each grid cell Gk, only the average depth or occupancy status is retained.

[0098] During rendering, only low-resolution blocky areas, blurred areas, or uniformly filled areas are output. Furthermore, coordinate quantization is performed on distal joints such as the wrist and ankle.

[0099] Coordinate quantization reduces the risk of inferring fine motor skills of the extremities. In some implementations, the bones between the wrist and elbow joints, and between the ankle and knee joints, may be displayed only as directional line segments, without showing the precise location of the end-joints.

[0100] Regarding output frame rate, Level 1 can maintain... Level 2 reduces the output frame rate to To further reduce the risk of continuous actions being reversed.

[0101] At Level 3, the graded desensitization rendering module 710 stops outputting continuous stick figure images, continuous stick figure video streams, continuous human body contours, high-frequency skeletal coordinates, fine joint coordinates, torso contours, and end joint trajectories. The system only outputs event-level data, such as: { event_type: "Sensitive nursing care action in progress", start_time: "Event start time", end_time: "Event end time", region_id: "Sensitive region ID", person_count: 2, privacy_level: 3, display_status: "Continuous attitude stream has been paused" } In Level 3 mode, the system can still process depth maps, infrared maps, radar point clouds, and attitude sequences for short periods of time to determine whether sensitive actions have ended, whether the human body has left the sensitive area, whether the degradation conditions are met, output the duration of the event, and restore Level 2 or Level 1 output, but it does not output the original or unsensitized data to the outside world.

[0102] S10. Output the anonymization result corresponding to the current privacy level. The data output module 730 outputs the anonymization result according to the current privacy level. For example... Figure 7 The diagram illustrates the stick figure generation interface of this invention. This interface displays anonymized two-person stick figure images generated by the stick figure generation module 720 based on skeletal topology. It can distinguish different human figures using different colors, line types, or numbers, and identify occluded or completed joints. Simultaneously, it presents corresponding desensitization effects according to changes in the current privacy level. Level 1 outputs anonymized two-person stick figure images, anonymized two-person stick figure video streams, torso and limb skeletons, major joint coordinates, human identification numbers, human speed, occluded joint list, completed joint list, action type, and timestamp. Level 2 outputs the stick figure image with the head hidden, torso centerline, regularized torso region, quantized end joint coordinates, down-frequency pose flow, action type, number of people, timestamp, and privacy level. Level 3 only outputs event type, event start time, event end time, event duration, number of people, coarse-grained region, privacy level, and continuous pose flow pause status. Through the complete technical chain described above, this invention achieves stable identity, accurate completion of occluded joints, and high-precision 3D skeletal modeling in two-person close care scenarios. At the same time, it achieves a good balance between regulatory continuity and privacy protection through a hierarchical privacy mechanism.

[0103] Based on the above method, the software system of the present invention may include modules as shown in Table 2, and the functional descriptions of each module are shown in Table 2: Table 2

[0104] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. A method for multimodal dual-person pose modeling and privacy-preserving stick figure generation, characterized in that, include: S1 acquires infrared structured light depth map, infrared map, radar point cloud and micro-Doppler data under minimized control at the acquisition end, and shuts down or isolates RGB image stream during normal operation; S2 performs filtering, back projection, and human instance segmentation on the structured light data to extract key points of the 3D skeleton of two people, joint confidence, and human mask. S3 performs filtering and static background removal on radar point clouds, performs dual-person clustering and trajectory tracking based on spatial position and velocity, and constructs gated regions for each target to extract micro-Doppler time-frequency features. S4 performs time synchronization and spatial registration, mapping the structured light skeleton, mask, radar point cloud, and trajectory to a unified world coordinate system; S5 performs bidirectional data-level fusion, using radar points to fill in structured light depth holes and using structured light human body masks to remove radar clutter points; S6. Construct an identity association cost function that includes centroid, velocity, size, historical trajectory and micro-Doppler differences, and determine the correspondence between structured light instances and radar trajectories by matching the minimum total cost. S7, for each target after identity association, introduces the structured light side and radar side feature inputs into a customized multimodal feature fusion model of skeleton topological position encoding, micro-Doppler additional channel, confidence mask and identity gating, and outputs a refined three-dimensional skeleton; The customized multimodal feature fusion model is constructed and run as follows: the skeleton of each human body is represented as a graph structure G=(V,E) consisting of joint nodes and bone connections, and a topological position code PEtopo is generated for each joint node based on joint type, left and right side identifiers and topological distance on the graph, in order to distinguish joints with different semantics but similar spatial appearance. On the structured light side, the joint features, the topological position encoding, and the joint confidence are concatenated to form a query vector Q; On the radar side, the micro-Doppler time-frequency characteristics within the gated area are incorporated into the key vector K and value vector V as an additional micro-Doppler channel to enhance the perception of human motion rhythm. Apply a joint confidence mask to joints with confidence scores below a threshold to increase the attention weight of the joint to the radar compensation branch, so that radar features can compensate for the absence or unreliability of structured light joints. An identity gating mask is applied to the cross-attention weights so that each structured light joint only focuses on the key value corresponding to the radar gating region that matches its identity, thereby suppressing the cross-contamination of radar features between two people. Residual connections are introduced during the fusion process to preserve the high-precision priors of the structured light joints and stabilize training. The multi-head cross-attention method is used to calculate and fuse attention in four subspaces: spatial position, velocity direction, micro-Doppler time-frequency texture, and historical trajectory. Finally, the regression head outputs the refined three-dimensional coordinates and fusion confidence of each joint, so that fine-grained joints such as wrist, elbow, knee and ankle can still achieve higher localization accuracy than general attention models under sparse radar point cloud conditions, and provide a reliable basis for subsequent occluded joint determination and completion. S8 determines occluded or missing joints based on confidence level, depth validity and historical trajectory, and completes them by combining radar velocity direction, micro-Doppler trend and kinematic constraints and adds completion labels. S9 determines the current privacy level based on sensitive areas, sensitive actions, duration of contact between two people, and external commands, using dynamic switching logic with hysteresis mechanism; S10, output according to the current privacy level: Level 1 outputs anonymized two stick figures, Level 2 hides the head, desensitizes the torso, quantifies the ends and reduces the frequency, Level 3 only outputs event-level results.

2. The method for multimodal dual-person pose modeling and privacy stick figure generation according to claim 1, characterized in that, The bidirectional data-level fusion described in S5 also includes: determining the structured light modal uncertainty based on the structured light depth map void ratio and human mask integrity. Radar mode uncertainty is determined based on the radar point cloud density and the mean signal-to-noise ratio. ,according to and The fusion weights are dynamically allocated and weighted according to the weights in deep hole completion and radar clutter filtering.

3. The method for multimodal dual-person pose modeling and privacy stick figure generation according to claim 1, characterized in that, The micro-Doppler difference described in S6 is obtained by extracting the length of each human body within the gated region. The sliding time window micro-Doppler time-frequency sequence is obtained by calculating its time-frequency texture similarity. When the length of historical data is less than the minimum effective window If the micro-Doppler energy is below the threshold, the weight of the micro-Doppler term is dynamically reduced or set to zero, degenerating into an association strategy that mainly relies on the centroid, size, and historical trajectory.

4. The method for multimodal dual-person pose modeling and privacy stick figure generation according to claim 1, characterized in that, The completion in S8 includes: when a joint meets the conditions of confidence level below a threshold, invalid corresponding depth, or missing multiple consecutive frames, the velocity direction of the historical trajectory extrapolation position is corrected by combining the radar local velocity direction vector and micro-Doppler motion trend. At the same time, human skeleton length constraints and left and right limb symmetry constraints are applied to generate the completed joint coordinates and completion confidence.

5. The method for multimodal dual-person pose modeling and privacy stick figure generation according to claim 1, characterized in that, The sensitive actions described in S9 are identified based on a continuous three-dimensional skeletal sequence within a time window, the distance between the centroids of two people, the limb contact relationship, and the area where the human body is located. When the confidence level of a sensitive action is not lower than the first threshold θs and the upgrade frame count is met consecutively, the privacy level is upgraded from Level 1 to Level 2; when the confidence level of a highly sensitive action is not lower than the second threshold θh and the upgrade frame count is met consecutively, the level is upgraded to Level 3.

6. The method for multimodal dual-person pose modeling and privacy stick figure generation according to claim 1, characterized in that, The Level 2 trunk desensitization includes at least one of the following methods: Method A: Calculate the midpoint of the shoulder and the midpoint of the hip, render only the center line segment between them, and hide the chest width, waist width and body curves; Method B replaces the quadrilateral formed by the left and right shoulders and the left and right hips with a standardized rectangle or trapezoid, and performs interval quantization on the shoulder width and hip width in steps of 5–15cm. Method C performs downsampling and curve simplification on the torso contour point set, retaining only the approximate shape of the area occupied by the human body; Method D divides the torso region into low-resolution grids, with each grid retaining only the average depth or occupancy state, and rendering it as a blocky or blurred area.

7. The method for multimodal dual-person pose modeling and privacy stick figure generation according to claim 1, characterized in that, The hysteresis mechanism is set as follows: Upgrading from Level 1 to Level 2 requires 10 consecutive frames. Downgrading from Level 2 to Level 1 requires 30 consecutive frames and a minimum holding time of 2 seconds; Upgrading from Level 2 to Level 3 requires 10 consecutive frames. Downgrading from Level 3 to Level 2 requires a continuous 60 frames and a minimum holding time of 5 seconds; The 10 frames, 30 frames, and 60 frames are fixed frame number thresholds; Under variable frame rate conditions, the fixed frame rate threshold is automatically converted into the corresponding time threshold.

8. A multimodal dual-person pose modeling and privacy-preserving stick figure generation system, used to implement the multimodal dual-person pose modeling and privacy-preserving stick figure generation method according to any one of claims 1-7, characterized in that, It includes a data acquisition module (210), a data minimization module (220), a structured light preprocessing module (310), a radar preprocessing module (320), a spatiotemporal alignment module (410), a bidirectional data-level fusion module (420), a dual-person identity association module (430), a customized multimodal feature fusion module (440), an occlusion joint intelligent completion module (450), a sensitive area and action recognition module, a dynamic privacy level determination module with hysteresis, a graded desensitization rendering module (710), and a stick figure and event-level output module.

9. A multimodal dual-person pose modeling and privacy stick figure generation system according to claim 8, characterized in that, The system also includes an infrared structured light device (110) and a 4D millimeter-wave radar (120). The infrared structured light device (110) is used to collect depth data of indoor scenes, and the 4D millimeter-wave radar (120) is used to collect human point clouds, radial velocity, signal-to-noise ratio and micro-Doppler slow time series. The data acquisition module (210) is used to acquire infrared structured light data and 4D millimeter-wave radar data; The data minimization module (220) at the acquisition end is used to turn off, filter, or isolate unnecessary RGB image streams; The structured light preprocessing module (310) is used to filter the depth map, segment human instances, extract skeletal key points, and extract human contours. The radar preprocessing module (320) is used to filter, perform two-person clustering, target tracking, and extract micro-Doppler features from the radar point cloud; The spatiotemporal alignment module (410) is used to complete the time synchronization and spatial coordinate unification of the two types of data; The bidirectional data-level fusion module (420) is used to complete the radar point cloud filling of structured light depth holes and structured light mask constraint of radar clutter; The dual-person identity association module (430) is used to perform dual-person identity matching based on location, speed, size, micro-Doppler features and historical trajectory; The customized multimodal feature fusion module (440) is used to output a refined three-dimensional skeleton through a multimodal fusion model; The occluded joint intelligent completion module (450) is used to intelligently complete low-confidence, missing or occluded joints; The hierarchical desensitization rendering module (710) is used to perform head hiding, torso simplification, contour quantization, raster blurring and end joint quantization according to the privacy level; The sensitive area and action recognition module includes a sensitive area recognition module (510) and a sensitive action recognition module (520); The sensitive area identification module (510) is used to determine whether a human body has entered a preset sensitive area; The sensitive action recognition module (520) is used to recognize sensitive actions based on continuous skeletal sequences, distance between two people, and limb contact relationships; The dynamic privacy level determination module with hysteresis includes a privacy level determination module (610) and a privacy level switching module (620); The privacy level determination module (610) is used to calculate the current privacy level based on the basic level, sensitive area, sensitive action, contact relationship and external instruction; The privacy level switching module (620) is used to perform privacy level upgrades and downgrades according to the hysteresis mechanism; The stick figure and event-level output module includes a stick figure generation module (720) and a data output module (730). The stick figure generation module (720) is used to generate a two-person stick figure image, video stream or event-level result based on the skeletal topology. The data output module (730) is used to output the desensitized result corresponding to the current privacy level.

Images