A reliable three-dimensional human pose reconstruction method based on millimeter-wave radar

By performing temporal alignment and compensation processing on radar point clouds, combined with human physical constraints and statistical uncertainty assessment, the problems of temporal misalignment and interpolation noise interference in radar point cloud data were solved, achieving high-precision and reliable 3D pose reconstruction.

CN123112902APending Publication Date: 2026-11-17海尔集团(青岛)金盈控股有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611431737.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies suffer from temporal misalignment, interpolation noise interference, and cross-modal distillation label errors in the 3D attitude reconstruction of radar point cloud data, resulting in low attitude reconstruction accuracy and a lack of reliability assessment.

Method used

By using time alignment and compensation processing, a compensation identifier is introduced to distinguish between real and compensated detection points. Combining human physical constraints and statistical uncertainty assessment, a multi-hypothesis generation and feature weight suppression mechanism is adopted to output joint-by-joint credibility.

Benefits of technology

It improves the accuracy and reliability of attitude reconstruction, ensuring that the output results are physically interpretable and reliable, and is suitable for high-reliability 3D attitude reconstruction in complex scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123112902A_ABST
    Figure CN123112902A_ABST
Patent Text Reader

Abstract

This application relates to the fields of radar perception and artificial intelligence technology, and provides a method for 3D reliable human posture reconstruction based on millimeter-wave radar. The method includes: acquiring a radar point cloud sequence and performing temporal alignment and compensation processing to obtain a target point cloud sequence with compensation labels; extracting features and generating multiple hypotheses from the target point cloud sequence using a posture reconstruction model to obtain multiple candidate postures, and suppressing the feature weights of compensation detection points based on compensation labels during feature extraction; performing consistency screening on the candidate postures based on human physical constraints to obtain a target posture set; determining the joint-by-joint reliability based on the statistical uncertainty of the target posture set and radar spatial observation conditions, and outputting the 3D posture estimation result. This application eliminates temporal misalignment and interpolation noise interference, achieving high-precision and physically interpretable 3D posture reconstruction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of radar perception and artificial intelligence technology, specifically to a method for reconstructing a three-dimensional reliable human posture based on millimeter-wave radar. Background Technology

[0002] In radar-based non-contact human posture perception technology, radar point cloud data is typically characterized by sparsity and intermittent missingness. Existing technologies often directly perform temporal modeling or simple interpolation on the original point cloud sequence when reconstructing 3D posture. However, due to the temporal discrepancy between the radar sampling time and the target's motion state, and the lack of real physical motion information in false detection points introduced by interpolation, temporal feature misalignment and noise amplification occur, severely impacting posture estimation accuracy. Furthermore, in the cross-modal knowledge distillation process using visual sensors to assist in training radar posture models, the visual teacher labels themselves contain depth errors and structural distortions. Directly using these as supervisory signals can lead to error propagation. Simultaneously, existing methods lack an effective mechanism for measuring the physical plausibility of the output posture, making it difficult to provide reliable confidence assessments in practical deployments. Therefore, there is an urgent need for a reconstruction method that can eliminate radar point cloud temporal misalignment and interpolation noise interference, block teacher label error propagation, and output a highly reliable 3D posture with physical interpretability. Summary of the Invention

[0003] To address the problems of low pose reconstruction accuracy and lack of credibility assessment caused by radar point cloud timing misalignment, interpolation noise interference, and cross-modal distillation tag noise in existing technologies, this application proposes a three-dimensional credible human pose reconstruction method based on millimeter-wave radar, which achieves high-precision and physically interpretable three-dimensional pose estimation and joint-by-joint credibility output.

[0004] To achieve the above objectives, this application adopts the following technical solution:

[0005] A radar point cloud sequence is acquired, and temporal alignment and compensation processing are performed on it to obtain a target point cloud sequence. Detection points in the target point cloud sequence are marked with compensation identifiers indicating the data source. Feature extraction and multiple hypothesis generation are performed on the target point cloud sequence using an attitude reconstruction model to obtain multiple candidate attitudes. During feature extraction, feature weight suppression is performed on detection points originating from the compensation processing based on the compensation identifiers. Consistency screening is performed on the multiple candidate attitudes based on human physical constraints to obtain a target attitude set. Based on the statistical uncertainty of the target attitude set and the spatial observation conditions of the radar on the target joints, the joint-by-joint confidence level is determined, and the 3D attitude estimation results and joint-by-joint confidence levels are output.

[0006] The above scheme reduces the interference of interpolation noise on model learning from the source by introducing compensation labels to distinguish between real detection points and compensated detection points in the preprocessing stage and specifically suppressing the feature weights of compensated data in subsequent feature extraction. At the same time, it combines human physical constraints to screen multiple hypotheses and integrates statistical uncertainty and radar spatial observation conditions to evaluate credibility, so that the output results are not only highly accurate, but also have clear physical interpretability and reliability boundaries.

[0007] As one implementation method, temporal alignment and compensation processing is performed on the radar point cloud sequence to obtain the target point cloud sequence, including: locating continuous empty frame segments in the radar point cloud sequence, performing interpolation compensation based on adjacent valid frames of the continuous empty frame segments to generate compensated detection points, and marking the compensation identifier of the compensated detection points as the interpolation compensation type; for the real detection points in the radar point cloud sequence, calculating the displacement compensation amount based on the radial Doppler velocity of the real detection points, superimposing the displacement compensation amount onto the original spatial coordinates of the real detection points to obtain the corrected spatial coordinates aligned to the target time; for the compensated detection points whose compensation identifier is the interpolation compensation type, keeping their interpolation coordinates unchanged; and constructing the target point cloud sequence based on the real detection points and the compensated detection points.

[0008] This implementation solves the motion blur problem caused by inconsistent sampling times by using the radar's unique radial Doppler velocity information to perform spatiotemporal extrapolation and alignment of the real detection points. At the same time, it skips the extrapolation operation for interpolation compensation points, avoiding the introduction of additional position errors in the absence of real velocity information, and ensuring the physical rationality of the timing alignment.

[0009] As one implementation method, interpolation compensation is performed based on adjacent valid frames of a continuous empty frame segment to generate compensation detection points. This includes: determining the preceding valid frame and the following valid frame in the continuous empty frame segment; using the spatial coordinates and radial Doppler velocity of each detection point in the preceding valid frame, propagating each detection point forward to the time of each empty frame in the continuous empty frame segment according to a uniform motion model to generate a forward compensation detection point; when the frame interval between the preceding valid frame and the following valid frame is greater than a preset single-frame interpolation length, using the spatial coordinates and radial Doppler velocity of each detection point in the following valid frame to perform interpolation compensation on the following frame. For forward propagation, the forward compensation detection points and backward propagation detection points are weighted and fused according to the time distance ratio of each empty frame relative to the forward and backward effective frames to obtain the compensation detection points. The displacement compensation amount is calculated based on the radial Doppler velocity of the real detection points, including: calculating the radial unit vector from the radar origin to the real detection points according to the spatial coordinates of the real detection points; taking the product of the radial Doppler velocity of the real detection points and the radial unit vector as the three-dimensional velocity vector; calculating the time difference between the current frame and the target time, and taking the product of the three-dimensional velocity vector and the time difference as the displacement compensation amount.

[0010] This implementation achieves a smoother interpolation transition in longer empty frame segments through a weighted fusion strategy of forward and backward effective frames, avoiding excessive cumulative errors in unidirectional propagation; at the same time, it accurately decomposes the one-dimensional radial Doppler velocity into three-dimensional spatial displacement, ensuring the geometric accuracy of extrapolation alignment in three-dimensional space.

[0011] As one implementation method, a pose reconstruction model is used to extract features and generate multiple hypotheses from the target point cloud sequence to obtain multiple candidate poses. This includes: in the spatial feature extraction stage, each detection point in each frame of the point cloud is mapped to a point-level feature vector, and a differentiated attention mask is constructed based on the compensation identifier, so that the attention weight of detection points not derived from the compensation processing is higher than the attention weight of detection points derived from the compensation processing. The point-level feature vectors are aggregated into a frame-level feature vector through mask attention pooling. In the temporal encoding stage, the frame-level feature vectors of each frame are arranged in chronological order. For frames containing detection points derived from the compensation processing, a position encoding attenuation factor is determined based on the proportion of detection points derived from the compensation processing to the total number of detection points in the frame. The position encoding attenuation factor is superimposed on the position encoding to reduce the contribution weight of the frame to the temporal context. Human condition features are extracted after self-attention encoding. Multiple candidate poses are generated based on the human condition features.

[0012] This implementation method suppresses the negative impact of compensation data in both spatial and temporal dimensions: spatially, it reduces the feature response intensity of interpolation points through attention masks, and temporally, it weakens the contextual relevance of high interpolation ratio frames through position encoding attenuation, thereby forcing the model to rely more on real observation data for pose inference and improving the robustness of the model.

[0013] As one implementation method, feature extraction is performed on the target point cloud sequence through a pose reconstruction model, including: multi-scale feature extraction of the point cloud in the target point cloud sequence to obtain spatial features at different scales; cross-scale fusion of spatial features at different scales to obtain fused spatial features; and subsequent temporal coding and pose generation based on the fused spatial features.

[0014] This implementation enhances the model's ability to perceive human structures in different spatial ranges through a multi-scale feature extraction and fusion mechanism. It retains the fine features of local joints while capturing the macroscopic structural information of the entire torso, providing a richer spatial representation foundation for subsequent temporal modeling and pose generation.

[0015] As one implementation method, a set of target poses is obtained by consistency screening of multiple candidate poses based on human physical constraints. This includes: for each candidate pose, calculating the relative deviation between the length of each bone segment and a preset human bone segment length reference value, and calculating the inter-frame joint velocity relative to the output pose of the previous frame; marking candidate poses with relative deviations exceeding a preset bone segment deviation threshold as bone segment anomaly hypotheses, and marking candidate poses with inter-frame joint velocities exceeding a preset velocity threshold as velocity anomaly hypotheses; removing candidate poses that are simultaneously marked as bone segment anomaly hypotheses and velocity anomaly hypotheses from multiple candidate poses to obtain the set of target poses; determining the joint-by-joint reliability based on the statistical uncertainty of the set of target poses and the spatial observation conditions of the target joints by radar, including: calculating the joint-by-joint coordinate variance of multiple candidate poses in the set of target poses as statistical uncertainty; calculating the spatial distance between each detection point and each joint in the target point cloud sequence based on the spatial position of each joint in the three-dimensional pose estimation results, and estimating the radar observability prior value of each joint based on the spatial distance; and weightedly fusing the statistical uncertainty and the radar observability prior value to obtain the joint-by-joint reliability.

[0016] This implementation adopts a tolerant screening strategy of "removing only those with two anomalies," which avoids discarding valid hypotheses due to misjudgment of a single constraint and improves the stability of the system in complex scenarios. At the same time, the confidence assessment integrates the statistical distribution characteristics of the model output with the physical observation conditions of the radar sensor, so that the confidence index can reflect both the uncertainty of the model itself and the actual observation quality of the sensor on a specific joint, thus achieving true physical interpretability confidence.

[0017] As one implementation method, candidate postures that are simultaneously labeled with both the bone segment abnormality hypothesis and the velocity abnormality hypothesis are removed from multiple candidate postures to obtain a target posture set. This further includes: if the removed candidate posture set is not empty, then the removed candidate posture set is used as the target posture set, and the joint-by-joint mean of the candidate postures in the target posture set is taken as the three-dimensional posture estimation result; if the removed candidate posture set is empty, then multiple candidate postures are used as the target posture set, and the joint-by-joint mean of the multiple candidate postures is taken as the three-dimensional posture estimation result; a weighted fusion of statistical uncertainty and radar observability prior values ​​is performed to obtain joint-by-joint reliability, including: when the radar observability prior value is lower than a preset observability threshold, the joint-by-joint reliability is suppressed to a preset reliability lower limit; when the radar observability prior value is not lower than the preset observability threshold, the statistical uncertainty and the radar observability prior value are weighted and summed according to a preset weight ratio to obtain the joint-by-joint reliability.

[0018] This implementation incorporates a fallback mechanism to ensure that even if all assumptions are deemed anomalous in extreme cases, a usable attitude estimation result can still be output, avoiding service interruption. Simultaneously, a lower confidence limit is forcibly applied to low-observable joints during confidence calculation, preventing the model from blindly outputting high-confidence results when radar observations are lacking, further enhancing security and reliability.

[0019] As one implementation method, the pose reconstruction model is trained as follows: Simultaneously acquire radar point cloud sample sequences and image sample sequences from an auxiliary vision sensor; generate initial teacher labels based on the image sample sequences using a pre-trained 3D human reconstruction model; calculate joint-by-joint soft confidence weights for each joint in the initial teacher labels based on depth effective range verification and inter-frame bone segment length consistency deviation, resulting in quality-weighted teacher labels; train the pose reconstruction model using the temporally aligned radar point cloud sample sequences as input and the quality-weighted teacher labels as the supervised target; during training, differentially weight the position loss according to the joint-by-joint soft confidence weights; in the multi-hypothesis generation module of the pose reconstruction model, a grouped annealing strategy is used to regularize the KL divergence loss: human joints are divided into trunk joint groups and limb distal joint groups; when calculating the reconstruction loss, a first annealing weight is applied to the trunk joint group, and a second annealing weight is applied to the limb distal joint group, with the growth rate of the second annealing weight being slower than that of the first annealing weight.

[0020] This implementation uses multidimensional physical constraints to quality-weight teacher labels, effectively filtering out depth errors and structural distortions in visual labels and purifying the supervision signal. At the same time, it adopts a grouped annealing strategy, which preserves greater diversity of posture assumptions of the limbs while ensuring trunk stability, enabling the model to better learn the natural variability of human movement and avoid overfitting to a single deterministic posture.

[0021] In addition, this application also provides an electronic device, including: a processor; a memory for storing a computer program; and a processor for executing the computer program to implement a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar as described above.

[0022] By executing the aforementioned method, the electronic device can complete the entire process of radar point cloud temporal alignment, feature suppression, multiple hypothesis generation, physical screening, and credibility assessment in real time, either locally or at the edge, meeting the requirements for low latency and high reliability in practical applications.

[0023] In addition, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar as described above.

[0024] The aforementioned storage medium enables the method to be distributed and deployed in software form, facilitating flexible integration and updates across different hardware platforms.

[0025] Beneficial effects:

[0026] This application presents a reliable 3D human pose reconstruction method based on millimeter-wave radar. By introducing a temporal extrapolation alignment mechanism based on radial Doppler velocity in the preprocessing stage, and combining it with compensation labels to implement spatial and temporal dual-weight suppression in the feature extraction stage, this fundamentally eliminates temporal misalignment and noise interference caused by sampling gaps and interpolation compensation in radar point clouds, significantly improving the quality of input data for pose reconstruction. Furthermore, it employs human physical constraints to tolerantly screen multiple hypotheses and integrates statistical uncertainty with radar spatial observation conditions to evaluate joint-by-joint reliability. This results in outputs that not only possess high accuracy but also clear physical interpretability and reliability boundaries, addressing the lack of reliability measurement in existing technologies. Simultaneously, during the training stage, multi-dimensional physical constraint-weighted teacher labels and a grouped annealing strategy effectively block the propagation path of visual label errors and preserve the natural variability of human motion, further enhancing the model's generalization ability and robustness. The overall scheme forms a complete closed loop from data preprocessing and model training to inference evaluation, applicable to high-reliability 3D pose reconstruction tasks in various radar perception scenarios. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar according to an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0030] Example 1

[0031] like Figure 1 As shown in the figure, this application provides a method for reconstructing a reliable three-dimensional human posture based on millimeter-wave radar. This method is mainly applied to radar sensing devices or edge computing terminals to recover high-precision three-dimensional human posture information from sparse and discontinuous radar echo data. The method includes the following steps S10 to S40.

[0032] Step S10: Obtain the radar point cloud sequence, perform temporal alignment and compensation processing on the radar point cloud sequence to obtain the target point cloud sequence; the detection points in the target point cloud sequence have compensation markers indicating the data source. Specifically, due to the limitations of the scanning mechanism and target reflection characteristics, the original point cloud output by the radar sensor is often non-uniformly sampled on the time axis, and there are empty frames due to target occlusion or weak signals. In order to construct a continuous data stream suitable for temporal modeling, this embodiment first performs temporal alignment and compensation processing on the original radar point cloud sequence. The "radar point cloud sequence" mentioned in this embodiment is only exemplified as a typical preferred form of radar spatial sensing data, and it is not a unique limitation on the input data format of this application. In practical applications and other feasible alternative implementations, radar sensing data can also exist in various other directly processable radar data structures or spatial tensor forms, such as radar multidimensional heat maps or radar target point array sequences after clustering or tracking filtering. Accordingly, the "detection point" mentioned in this application should be broadly understood as the basic spatial sensing unit in the radar sensing data. When the input data is a discrete point cloud, the "detection point" corresponds to a spatial coordinate point; when the input data is a radar heat map, feature map, or three-dimensional voxel grid, the "detection point" is equivalent to a pixel, spatial grid cell, or voxel cell whose energy response value in the heat map, feature map, or voxel meets a preset threshold.

[0033] Temporal alignment refers to mapping detection points collected at different times to a unified time base to eliminate motion blur caused by differences in sampling times. Compensation processing, on the other hand, addresses missing data periods by generating virtual detection points using information from neighboring valid frames to maintain temporal continuity. Crucially, a compensation identifier is attached to each detection point during processing. This identifier serves as metadata throughout the entire process, explicitly indicating whether the detection point originates from real echoes directly detected by radar or from compensation data generated by the algorithm. For example, the compensation identifier can be embedded in the point cloud data structure in the form of a Boolean type, an enumeration type, or a confidence score. This explicit source marking mechanism enables subsequent processing modules to distinguish between actual observation information and speculative completion information, providing a data foundation for suppressing interpolation noise.

[0034] Step S20 involves feature extraction and multiple hypothesis generation of the target point cloud sequence using a pose reconstruction model to obtain multiple candidate poses. During feature extraction, feature weight suppression is performed on the detection points derived from the compensation processing based on compensation identifiers. Specifically, the pose reconstruction model can be a nonlinear mapping model built based on deep learning, graph neural networks, or other machine learning paradigms. Its input is the target point cloud sequence, and its output is a set of candidate poses representing the spatial positions of human joints. Due to the sparsity and multipath effect of radar point clouds, a single deterministic pose estimation is prone to getting trapped in local optima or producing artifacts. Therefore, this embodiment adopts a multiple hypothesis generation strategy, that is, the model outputs multiple possible pose solutions simultaneously to cover the uncertainty distribution of the true pose. In the feature extraction stage, to avoid the spurious geometric structures introduced by the compensation detection points interfering with model learning, this embodiment introduces a feature weight suppression mechanism based on compensation identifiers. Here, feature weight suppression is a functional definition, referring to any technical means that can reduce the contribution of compensation detection points to the internal feature representation of the model. For example, a differential attention mask can be constructed to make the attention weight of real detection points significantly higher than that of compensated detection points; or the weighting coefficients of each point can be dynamically adjusted according to the compensation identifier during feature aggregation; or a penalty term can be applied to the feature responses originating from the compensation points in the loss function calculation. Through this suppression mechanism, the model is guided to pay more attention to the highly reliable real observation data, thereby effectively reducing the noise accumulation effect caused by interpolation compensation while ensuring temporal continuity.

[0035] Step S30 involves consistency screening of multiple candidate postures based on human physical constraints to obtain a target posture set. Specifically, while the multiple candidate postures generated by the model cover the possible solution space, they may contain abnormal solutions that violate the laws of human biomechanics. Therefore, this embodiment introduces human physical constraints as a post-processing screening criterion. Human physical constraints should be interpreted broadly, including not only the constant constraint on the length of bone segments between adjacent joints and the continuity constraint on joint movement speed, but also constraints on the range of joint motion angles, limb symmetry, and trunk rigidity. During the screening process, each candidate posture is checked to see if it meets the preset physical constraint conditions, and candidate postures that meet the constraints or whose deviations are within the tolerance threshold are retained to form the target posture set. For example, if the femur length of a candidate posture exceeds 20% of the normal human anatomical range, and the instantaneous speed of the knee joint exceeds the human movement limit, then the posture will be judged as abnormal and eliminated. This screening mechanism based on prior knowledge is equivalent to adding a physical safety valve to the model output, effectively filtering out non-physiological postures caused by radar clutter or model illusions, significantly improving the biological rationality of the output results.

[0036] Step S40: Based on the statistical uncertainty of the target attitude set and the spatial observation conditions of the radar on the target joints, determine the joint-by-joint reliability and output the three-dimensional attitude estimation result and the joint-by-joint reliability. Specifically, traditional attitude estimation methods usually only output a single coordinate value and lack reliability measurement, which is a significant hidden danger in security-sensitive scenarios. This embodiment innovatively proposes a joint-by-joint granularity reliability evaluation mechanism, which integrates two dimensions: the statistical uncertainty within the model and the spatial observation conditions outside the sensor. The statistical uncertainty reflects the model's own confidence in the current attitude, which can be quantified, for example, by calculating the variance, entropy, or dispersion of the multimodal distribution of the same joint coordinates in the target attitude set. The spatial observation conditions reflect the actual detection quality of the radar sensor on the joint, for example, by comprehensively evaluating the prior value of radar observability based on factors such as the distance between the joint and the radar, the relative azimuth angle, the presence of obstruction, and the signal-to-noise ratio. Finally, the information from these two dimensions is weighted and fused or nonlinearly mapped to generate an independent reliability score for each joint. For example, even if multiple candidate poses output by the model are highly consistent (low statistical uncertainty), the final credibility will still be suppressed if the joint is located at the edge of the radar detection blind zone (poor spatial observation conditions). This two-factor fusion credibility assessment avoids blind confidence in the model and compensates for the shortcomings of relying solely on sensor indicators, providing users with physically interpretable decision-making basis. The final output of the 3D pose estimation result can be the mean, median, or maximum likelihood estimate of the target pose set, and is output together with the corresponding joint-by-joint credibility for downstream applications to perform graded response or risk warning.

[0037] Through steps S10 to S40 described above, this embodiment constructs a complete closed-loop processing flow from data preprocessing, feature suppression, multiple hypothesis generation to physical screening and credibility assessment. The core of this flow lies in fully utilizing the radar's unique velocity information and compensation identifiers, actively suppressing interpolation noise while preserving temporal continuity, and evaluating output reliability through both internal and external factors. This fundamentally solves the problems of low radar attitude reconstruction accuracy and lack of credibility measurement in existing technologies.

[0038] Example 2

[0039] Building upon Example 1, this example further details the specific implementation of temporal alignment and compensation processing. As one implementation method, temporal alignment and compensation processing is performed on a radar point cloud sequence to obtain a target point cloud sequence. This includes: locating consecutive empty frame segments in the radar point cloud sequence; performing interpolation compensation based on adjacent valid frames of the consecutive empty frame segments to generate compensated detection points; and marking the compensation identifier of the compensated detection points as the interpolation compensation type. For real detection points in the radar point cloud sequence, the displacement compensation amount is calculated based on the radial Doppler velocity of the real detection points; the displacement compensation amount is superimposed on the original spatial coordinates of the real detection points to obtain corrected spatial coordinates aligned to the target time. For compensated detection points with the compensation identifier as the interpolation compensation type, their interpolation coordinates remain unchanged. The target point cloud sequence is constructed based on the real detection points and the compensated detection points. Specifically, this processing method establishes a differentiated processing branch based on the data source. For real detection points, the Doppler effect information unique to radar is used for physical-level spatiotemporal correction to eliminate motion blur caused by sampling time deviations. For compensated detection points generated due to missing data, since they lack real instantaneous velocity measurements, their interpolated static coordinates are forcibly maintained to avoid introducing false motion vectors that could lead to secondary errors. This differentiated processing mechanism ensures that the target point cloud sequence maintains temporal continuity while preserving the physical authenticity of the actual observation data to the greatest extent possible, and provides an accurate data foundation for weight suppression based on compensation labels in the subsequent feature extraction stage.

[0040] Regarding the specific generation logic of interpolation compensation, as one implementation method, interpolation compensation is performed based on adjacent valid frames of a continuous empty frame segment to generate compensation detection points. This includes: determining the forward valid frame before the continuous empty frame segment and the backward valid frame after it; using the spatial coordinates and radial Doppler velocity of each detection point in the forward valid frame, propagating each detection point forward to the time of each empty frame in the continuous empty frame segment according to a uniform motion model to generate forward compensation detection points; when the frame interval between the forward valid frame and the backward valid frame is greater than the preset single-frame interpolation length, using the spatial coordinates and radial Doppler velocity of each detection point in the backward valid frame, performing backward propagation, and weighting and fusing the forward compensation detection points and the backward propagation detection points according to the time distance ratio of each empty frame relative to the forward valid frame and the backward valid frame to obtain compensation detection points. Here, the forward valid frame refers to the most recent frame of radar data located before the continuous empty frame segment on the time axis and containing at least one valid detection point, and the backward valid frame refers to the most recent frame of radar data located after the continuous empty frame segment and containing at least one valid detection point. The preset single-frame interpolation length is an empirical threshold set based on the radar frame rate and typical human movement speed. It is used to determine whether unidirectional linear extrapolation will produce unacceptable cumulative errors. When the empty frame segment is short, the target trajectory can be approximately reconstructed by simply extrapolating the forward effective frames at a constant speed. However, when the empty frame segment is long, the error from unidirectional propagation increases linearly with time. In this case, introducing a backward effective frame for bidirectional constraint becomes particularly necessary. The specific calculation method for time-distance weighted fusion can be: Let the time difference between a certain empty frame moment and the forward effective frame be... The time difference with the subsequent effective frame is Then the final coordinates of a detection point in the empty frame ,in The coordinates are obtained from forward propagation. The coordinates are obtained through backpropagation. This weighted strategy ensures that the closer an interpolation point is to a valid frame on a certain side, the greater its influence on that frame. This achieves a smooth transition that conforms to the motion trend within long-range frames, effectively avoiding trajectory abrupt changes or jitter that may be caused by simple linear interpolation.

[0041] For the extrapolation alignment calculation of the actual detection points, one implementation method involves calculating the displacement compensation based on the radial Doppler velocity of the actual detection points. This includes: calculating the radial unit vector from the radar origin to the actual detection points based on their spatial coordinates; using the product of the radial Doppler velocity and the radial unit vector as the three-dimensional velocity vector; and calculating the time difference between the current frame and the target's moment, using the product of the three-dimensional velocity vector and the time difference as the displacement compensation. Specifically, since the Doppler velocity directly output by the millimeter-wave radar is only the radial component of the target relative to the radar line of sight, it cannot directly represent the full velocity vector in three-dimensional space. Therefore, it needs to be reconstructed through geometric relationships. Let the original spatial coordinates of the actual detection points be (x, y, z), and the radar origin be (0, 0, 0), then the radial unit vector... It can be represented as If the radial Doppler velocity at that point is The synthesized three-dimensional velocity vector This conversion is based on the geometric prior of radar detection, which assumes that the target's motion direction is mainly radial or its radial component dominates within a short period of time. This has high accuracy in scenarios where a human body is moving towards or away from the radar. Subsequently, let the current frame acquisition time be... The target alignment time is Then the time difference Displacement compensation amount The final corrected spatial coordinates are the vector sum of the original coordinates and the displacement compensation. Through the above vector decomposition and synthesis operations, the one-dimensional Doppler observations are accurately transformed into position correction quantities in three-dimensional space, enabling real detection points from different times to be uniformly mapped to the same time reference. This fundamentally solves the problems of point cloud trailing and loose structure caused by the mismatch between radar scanning cycle and target motion, and significantly improves the spatiotemporal consistency of the input data for subsequent attitude reconstruction models.

[0042] Example 3

[0043] Based on step S20 of Embodiment 1, this embodiment further details the dual noise suppression mechanism in the feature extraction process. As one implementation method, feature extraction and multiple hypothesis generation are performed on the target point cloud sequence using a pose reconstruction model to obtain multiple candidate poses. This includes: in the spatial feature extraction stage, each detection point in each frame of the point cloud is mapped to a point-level feature vector, and a differentiated attention mask is constructed based on the compensation identifier, making the attention weight of detection points not originating from the compensation process higher than the attention weight of detection points originating from the compensation process. The point-level feature vectors are aggregated into a frame-level feature vector through mask attention pooling. In the temporal coding stage, the frame-level feature vectors of each frame are arranged in chronological order. For frames containing detection points originating from the compensation process, a position coding attenuation factor is determined based on the proportion of detection points originating from the compensation process to the total number of detection points in that frame. The position coding attenuation factor is superimposed on the position coding to reduce the contribution weight of that frame to the temporal context. Human conditional features are extracted after self-attention coding. Multiple candidate poses are generated based on the human conditional features.

[0044] Specifically, the aforementioned feature extraction process constructs a dual noise suppression defense working collaboratively in both spatial and temporal dimensions. Its core lies in dynamically adjusting the model's attention to the input data using the data source confidence information carried by the compensation label, rather than employing a one-size-fits-all static threshold filtering. In the spatial feature extraction stage, the original attributes of each detection point, such as its 3D coordinates and reflection intensity, are first mapped into a high-dimensional point-level feature vector using a multilayer perceptron or linear layer. Subsequently, a corresponding attention mask value is generated based on the compensation label attached to each detection point. For example, for real detection points, the mask value can be set to a high-weight coefficient of 1.0 or close to 1.0; for compensated detection points, the mask value is set to a low-weight coefficient less than 1.0, such as a value between 0.1 and 0.5, or dynamically calculated based on the interpolation confidence. During attention pooling, this mask value directly affects the calculation of the attention score, allowing real detection points to dominate local feature aggregation, while compensated detection points only participate in the calculation as auxiliary geometric filling information. This mask attention pooling mechanism effectively avoids local geometric feature distortion caused by interpolation point position deviations or false distributions, ensuring that frame-level feature vectors can accurately represent the real human limb structure, rather than being misled by artificially generated smooth trajectories.

[0045] In the temporal coding stage, the dual suppression mechanism is further extended to the time dimension. The frame-level feature vectors of each frame are arranged in the order of acquisition time to form a temporal sequence, and a positional code is assigned to each frame to represent its temporal order. For frames containing compensated detection points, the proportion of compensated detection points in that frame to the total number of detection points is calculated in real time and denoted as the compensation ratio. Based on this compensation ratio, the positional coding attenuation factor is calculated, for example, using a linear attenuation formula. or exponential decay formula ,in This represents the position encoding attenuation factor of frame t. This indicates the compensation percentage for that frame. and The pre-defined attenuation control parameters are used. The calculated attenuation factor is superimposed or multiplied into the standard positional encoding, thereby reducing the temporal correlation strength of that frame in the subsequent self-attention encoding process. The physical significance of this design is that when a frame is mainly composed of interpolated data, it lacks real high-frequency motion details and transient change information. If it is given the same temporal weight as the real observed frame, it is very easy to cause the model to produce an over-smoothing effect or motion lag in temporal inference. By reducing the temporal contribution weight of low-quality frames, the model is guided to rely more on the adjacent real observed frames to infer intermediate states, thereby preserving the dynamic characteristics of real motion to the greatest extent while ensuring temporal continuity.

[0046] It should be understood that the specific calculation forms of the spatial attention mask and temporal position encoding attenuation factor are not limited to the examples listed in this embodiment. In other feasible implementations, spatial dimension weight suppression can also be achieved by multiplying the features of the compensation point by a scaling factor before feature concatenation, or by applying a penalty term to the feature responses originating from the compensation point in the loss function; temporal dimension attenuation can also be expressed as masking or scaling the attention scores of low-quality frames in the self-attention matrix. Regardless of the specific mathematical form used, as long as its essence is to distinguish the data source based on the compensation identifier and adjust the model's dependence on real observation data and compensation inferred data accordingly, it falls within the technical scope of this application. This dual suppression mechanism enables the attitude reconstruction model to adaptively balance the contradiction between temporal continuity and data authenticity under harsh data conditions such as sparse radar point clouds and frequent empty frames, significantly improving the robustness and physical reliability of the 3D attitude estimation results in complex scenes.

[0047] Example 4

[0048] Based on step S20 of Example 1, this example further details the specific architecture of spatial feature extraction in the pose reconstruction model. As one implementation method, feature extraction of the target point cloud sequence using the pose reconstruction model includes: multi-scale feature extraction of the point cloud in the target point cloud sequence to obtain spatial features at different scales; cross-scale fusion of spatial features at different scales to obtain fused spatial features; and subsequent temporal coding and pose generation based on the fused spatial features.

[0049] Specifically, given the significant spatial non-uniformity and sparsity of radar point cloud data, single-scale feature extraction often struggles to accurately represent both the macroscopic skeletal structure and microscopic limb details of the human body. For example, when the target is close to the radar, the point cloud density is high, and subtle changes in the posture of the hands or head contain rich geometric information; however, when the target is far away or during empty frame compensation periods, the point cloud is extremely sparse, only reflecting the general outline of the torso. If only a single receptive field network layer is relied upon, problems such as the smooth loss of local details or the fragmentation of the global structure due to noise are easily encountered. Therefore, this embodiment introduces a multi-scale feature extraction mechanism to construct a feature representation system that can adaptively perceive information at different spatial ranges. In this system, spatial features at different scales refer to a set of feature representations obtained by abstracting and aggregating point cloud data at different levels. High-resolution scale features retain more original geometric details and local neighborhood information, suitable for locating fine joint points; low-resolution scale features have undergone multiple downsampling or large receptive field aggregation, containing stronger semantics of human topology and global contextual information, which helps maintain the overall coherence of the skeleton when data is missing.

[0050] Regarding the specific implementation architecture for multi-scale feature extraction and cross-scale fusion, this application is not limited to a specific network topology, but covers any structure capable of achieving multi-level feature extraction and information interaction. As a preferred implementation, a hierarchical point cloud encoder combined with a Feature Pyramid Network (FPN) architecture can be adopted. In this architecture, the input target point cloud sequence is first downsampled level by level through pointwise convolutions or ensemble abstraction layers with shared weights, generating a series of hierarchical feature maps with decreasing resolution and increasing semantics. Subsequently, a top-down path is used to restore the high-level global semantic features to the same spatial resolution as the low-level features through upsampling operations, and lateral connections are used to concatenate or add the upsampled features with the original features of the corresponding level, thereby achieving cross-scale fusion. This fusion strategy ensures that the final fused spatial features contain both strong semantic guidance from deep networks and retain the fine spatial localization capabilities from shallow networks. It should be understood that in other feasible implementations, a U-Net-type symmetric encoder-decoder structure, a high-resolution network (HRNet), or a multi-stream parallel processing architecture can also be used to achieve the above functions. For example, in a U-Net variant, features from each stage of the encoder can be directly passed to the corresponding layer of the decoder for fusion via skip connections; in a multi-stream architecture, multiple parallel branches can be designed to process point cloud data of different voxel sizes or sphere query radii, and finally integrated along the feature dimension. Regardless of the specific network form used, as long as its core logic is to acquire and fuse multi-level spatial representations to enhance the model's adaptability to the sparsity of radar point clouds, it falls within the protection scope of this application.

[0051] After cross-scale fusion, the resulting fused spatial features are used as input to the subsequent temporal coding module. Specifically, the fused spatial features are typically represented as a sequence of high-dimensional feature vectors corresponding to each detection point or local region. This sequence not only carries rich spatial information enhanced by multiple scales within the current frame but also maintains its spatial correspondence with the original point cloud or regularized grid, enabling alignment and interaction with the differential attention mask and positional coding attenuation factor described in Example 3. In this way, multi-scale feature extraction is no longer an isolated preprocessing step but is deeply embedded in the inference chain of the entire pose reconstruction model. Its technical effects are twofold: firstly, it significantly improves the model's accuracy in capturing small-scale targets such as human end joints (e.g., wrists, ankles), reducing limb truncation or drift caused by sparse point clouds; secondly, guided by global structural features, it enhances the model's ability to infer the rationality of the overall pose in scenarios with severe occlusion or long-term interpolation, avoiding global deformation caused by local noise. This coarse-to-fine, multi-source fusion feature representation method lays a solid spatial representation foundation for the subsequent generation of high-quality, physically reliable multi-hypothetical poses.

[0052] Example 5

[0053] Based on steps S30 and S40 of Embodiment 1, this embodiment further details the specific implementation logic of physical constraint screening and multidimensional credibility assessment. As one implementation method, consistency screening of multiple candidate poses based on human physical constraints yields a target pose set, including: for each candidate pose, calculating the relative deviation between the length of each bone segment and a preset human bone segment length reference value, and calculating its inter-frame joint velocity relative to the previous frame's output pose; marking candidate poses with relative deviations exceeding a preset bone segment deviation threshold as bone segment anomaly hypotheses, and marking candidate poses with inter-frame joint velocities exceeding a preset velocity threshold as velocity anomaly hypotheses; removing candidate poses simultaneously marked as bone segment anomaly hypotheses and velocity anomaly hypotheses from multiple candidate poses to obtain the target pose set. Specifically, this screening mechanism employs a strategy combining double verification and tolerant removal. When calculating the relative deviation, each bone segment (such as the upper arm, forearm, thigh, lower leg, etc.) in the candidate pose is traversed, and its Euclidean distance length is compared with a preset human anatomical standard bone length or an adaptively updated personalized bone length reference value, for example, calculating... As a relative deviation, among The length of the bone segment in the candidate pose. The corresponding reference length is used. When calculating inter-frame joint velocities, the 3D displacement vector magnitude of the corresponding joint point is calculated using the current candidate pose and the final pose estimation result from the previous time step, and then divided by the frame interval. The logic of "simultaneously labeled" for elimination is adopted because the sparsity and multipath effect of radar point clouds may lead to instantaneous anomalies in a single dimension: for example, when a limb swings rapidly, Doppler blur may cause the velocity estimation to momentarily exceed the limit, but the bone length structure may still be correct at this time; conversely, when the human body is in an extreme twisting posture or there is occlusion, the bone length projection may be distorted, but the movement speed is still within a reasonable range. Only when a candidate pose violates both static anatomical constraints and dynamic kinematic constraints is it highly likely to be an invalid artifact or severe noise interference generated by the model. This "AND" logic elimination strategy effectively avoids the false negative elimination of valid hypotheses due to excessive sensitivity of a single indicator, and significantly improves robustness in complex dynamic scenes.

[0054] Regarding the output and reliability assessment of the filtered results, as one implementation method, candidate poses simultaneously marked as having bone segment abnormality and velocity abnormality hypotheses are removed from multiple candidate poses to obtain the target pose set. This further includes: if the removed candidate pose set is not empty, then the removed candidate pose set is used as the target pose set, and the joint-by-joint mean of the candidate poses in the target pose set is taken as the 3D pose estimation result; if the removed candidate pose set is empty, then multiple candidate poses are used as the target pose set, and the joint-by-joint mean of the multiple candidate poses is taken as the 3D pose estimation result. This fallback mechanism is designed to ensure continuous availability. In actual deployment, even if all candidate poses are judged as abnormal, no null value will be output or an error will be reported and interrupted; instead, the system will revert to the original set of multiple hypotheses and take the average. Although the pose accuracy may decrease at this point, basic human position and contour tracking can still be maintained, providing degraded services to downstream applications rather than a service black hole. This is crucial for scenarios such as security monitoring or elderly care where blind spots are unacceptable.

[0055] Based on this, as one implementation method, joint-by-joint confidence is determined based on the statistical uncertainty of the target attitude set and the spatial observation conditions of the target joints by the radar. This includes: calculating the joint-by-joint coordinate variance of multiple candidate attitudes in the target attitude set as the statistical uncertainty; calculating the spatial distance between each detection point and each joint in the target point cloud sequence based on the spatial position of each joint in the 3D attitude estimation results, and estimating the radar observability prior value of each joint based on the spatial distance; and weightedly fusing the statistical uncertainty and the radar observability prior value to obtain the joint-by-joint confidence. Specifically, the joint-by-joint coordinate variance reflects the degree of cognitive disagreement within the model regarding the current attitude solution: if the 3D coordinate distribution of a certain joint in the target attitude set is compact and the variance is extremely small, it indicates that multiple assumptions are highly consistent, and the model has high confidence in the positioning of that joint; conversely, if the variance is large, it indicates that the model has ambiguity or uncertainty at that point. However, relying solely on the statistical characteristics within the model is insufficient, as the model may produce a "consistency illusion" in areas lacking data support. Therefore, this embodiment introduces the radar observability prior value as an external physical correction factor. This prior value is negatively correlated with the spatial distance from the joint to the radar; for example, an inverse proportional function can be used. Alternatively, the Gaussian decay function can be used for calculation, where Let be the observability prior value of the k-th joint. This is the Euclidean distance from the joint to the radar origin. This represents the effective detection radius or normalization constant of the radar. Smaller spatial distances generally mean higher point cloud density and better signal-to-noise ratio at that joint, resulting in better radar observation quality and higher prior values. By weighted summing or multiplying the statistical uncertainty (usually converted to confidence level form, such as the inverse of variance or a negative exponential mapping) with the radar observability prior values, the resulting joint-by-joint confidence level encompasses both the model's own inference confidence and the verification of the sensor's physical observation capabilities, thus achieving a truly physically interpretable confidence assessment.

[0056] Furthermore, to prevent misleadingly high confidence levels from being output under extremely poor observation conditions, as one implementation method, statistical uncertainty and radar observability prior values ​​are weighted and fused to obtain joint-by-joint confidence. This includes: when the radar observability prior value is lower than a preset observability threshold, the joint-by-joint confidence is suppressed to a preset lower confidence limit; when the radar observability prior value is not lower than the preset observability threshold, the statistical uncertainty and the radar observability prior value are weighted and summed according to a preset weight ratio to obtain the joint-by-joint confidence. Specifically, the preset observability threshold is a safety red line set based on the actual performance boundary of the radar sensor, for example, corresponding to 80% of the radar's maximum effective detection range or the distance at which the signal-to-noise ratio drops to a certain critical value. When a joint is outside this threshold, regardless of how consistent the candidate attitudes output by the model are (i.e., how low the statistical uncertainty is), the confidence of that joint is forcibly lowered to the preset lower confidence limit (e.g., 0.1 or 0.2). This low-confidence suppression mechanism fundamentally prevents the model from exhibiting "blind confidence" in radar blind zones or weak signal areas, ensuring that the output credibility index always honestly reflects the true perceived quality. For example, in a smart home scenario for elderly fall detection, if a falling posture is detected but the head joint is suppressed to low confidence due to excessive distance, the downstream decision module can trigger a secondary review or lower the alarm priority, rather than directly issuing a false alarm, thereby significantly improving the safety and user experience of practical applications. It should be understood that the specific mathematical form, threshold selection, and suppression strategy of the above weighted fusion can all be adjusted according to the specific radar model and application requirements. As long as the core logic is to fuse the internal statistical characteristics of the model with the external observation conditions of the sensor to evaluate the reliability of each joint, it falls within the scope of protection of this application.

[0057] Example 6

[0058] This embodiment is specifically used to illustrate the method of obtaining the pose reconstruction model in the aforementioned embodiments. It should be noted that the training process described below is only performed during the model development and offline optimization phases, and these steps are not included in the actual deployment or inference phases. This description method, which physically separates the training process from the inference process, aims to clearly define the manufacturing and usage attributes of the product.

[0059] As one implementation method, the pose reconstruction model is trained as follows: radar point cloud sample sequences and image sample sequences from an auxiliary vision sensor are acquired simultaneously. A pre-trained 3D human reconstruction model is then used to generate initial teacher labels based on the image sample sequences. Specifically, simultaneous acquisition refers to ensuring, through hardware triggering or a high-precision timestamp alignment mechanism in the data acquisition system, that each frame of radar point cloud and its corresponding visual image are strictly matched in the temporal dimension. This is a prerequisite for cross-modal knowledge distillation. The auxiliary vision sensor can be an RGB-D camera, a binocular stereo camera, or other imaging devices capable of providing rich texture and depth information. The pre-trained 3D human reconstruction model can be a mature vision algorithm built based on existing technologies such as SMPL, HMR, or OpenPose. Its role is to utilize the high-resolution characteristics of visual information to generate initial teacher labels containing complete human joint coordinates. However, due to depth noise, occlusion blind spots, and domain differences inherent in vision sensors, the directly generated initial labels often contain non-negligible physical errors. If used indiscriminately for supervising radar model training, these errors can easily propagate and amplify.

[0060] To address the aforementioned issues, as one implementation method, based on depth effective range verification and inter-frame bone segment length consistency deviation, a joint-by-joint soft confidence weight is calculated for each joint in the initial teacher label, resulting in a quality-weighted teacher label. Specifically, this step constructs a multi-dimensional label quality evaluation system containing three core factors. The first factor is the depth effectiveness factor, which reads the depth values ​​of the depth image output by the auxiliary vision sensor at the projection positions of each joint in the initial teacher label and determines whether the depth value falls within the effective range of the sensor; if it exceeds the range or is in a high-noise edge region, a lower depth effectiveness factor is assigned, and vice versa. The second factor is the bone segment consistency factor, which calculates the rate of change of bone segment length between the current frame and the previous frame and compares this rate of change with a preset upper limit for human bone segment length change; since bone length should remain constant in real human movement, excessive inter-frame bone length fluctuations usually indicate structural distortion or tracking drift in visual reconstruction, therefore, the greater the rate of change, the lower the bone segment consistency factor. The third factor is the model confidence factor, which is the joint-by-joint prediction probability or confidence score output by the pre-trained 3D human reconstruction model itself, reflecting the degree of self-confidence of the visual algorithm in locating that joint. Finally, these three factors are multiplied and normalized to obtain an independent joint-by-joint soft confidence weight for each joint. This product-fusion mechanism ensures that a joint's label is only assigned a high weight when the depth is reliable, the structure is stable, and the model is confident; any anomaly in any dimension will lead to a significant decrease in weight, thus purifying the supervision signal at its source and blocking the transmission path of visual labeling errors to the radar pose reconstruction model.

[0061] After obtaining high-quality supervised targets, as one implementation method, a pose reconstruction model is trained using a temporally aligned radar point cloud sample sequence as input and quality-weighted teacher labels as supervised targets. During training, the position loss is differentially weighted according to the joint-by-joint soft confidence weights. Specifically, the temporally aligned radar point cloud sample sequence here refers to the training data processed by the same preprocessing procedures (including interpolation compensation and Doppler extrapolation alignment) described in Examples 1 to 3, to ensure the consistency of data distribution between the training state and the inference state. When calculating the position loss (such as L1 loss or L2 loss), all joints are no longer treated equally; instead, the position error of each joint is multiplied by the corresponding joint-by-joint soft confidence weight and then accumulated. This means that for joints with high label quality (such as the trunk center and clear close-range observation points), the model is driven by strong supervised signals to achieve accurate fitting; while for joints with low label quality (such as blurred distant points, occluded parts, or visual reconstruction anomalies), their loss contribution is automatically suppressed, avoiding the model being forced to learn incorrect geometric structures. This differentiated weighting strategy makes the training process more robust and significantly improves the model's generalization ability in complex scenarios.

[0062] Furthermore, to improve the quality of multi-hypothesis generation, as one implementation method, a grouped annealing strategy is used to regularize the KL divergence loss in the multi-hypothesis generation module of the pose reconstruction model: human joints are divided into trunk joints and limb distal joints. When calculating the reconstruction loss, the trunk joints are given a first annealing weight, and the limb distal joints are given a second annealing weight. The growth rate of the second annealing weight is slower than that of the first annealing weight. Specifically, in the early stages of training of a multi-hypothesis generation model (such as a conditional variational autoencoder CVAE), the KL divergence term often dominates the optimization direction, easily causing the posterior distribution to collapse prematurely to the prior distribution, resulting in multiple generated pose hypotheses tending towards a single mean and losing the ability to capture the diversity of human motion. This phenomenon is particularly severe in the distal extremities because the joints of the hands, feet, and other distal joints have large degrees of freedom of motion and sparse radar echoes, making them difficult to model accurately. Therefore, this embodiment introduces a grouped annealing strategy. The trunk joint group, including relatively stable core nodes such as the head, neck, shoulder, and hip, exhibits relatively regular movement patterns and has sufficient data support. Therefore, a rapidly increasing first annealing weight is used to quickly converge to a reasonable latent variable distribution, ensuring the stability of the overall skeleton. The limb distal joint group, including flexible nodes such as the elbow, wrist, knee, and ankle, employs a slower-increasing second annealing weight. This means that a small KL penalty is applied in the early stages of training, allowing the posterior distribution to maintain a large variance and exploration space. As training progresses, the second annealing weight increases slowly, gradually introducing regularization constraints. This differentiated annealing curve design allows the model to retain the diversity and multimodal characteristics of limb distal posture assumptions to the greatest extent while ensuring the correct trunk structure. It avoids the problems of distal posture smoothing or truncation caused by excessive regularization, thus truly achieving high-reliability, multimodal 3D posture reconstruction capabilities. It should be understood that the above grouping method, the specific functional form of the annealing weight (such as linear, exponential or stepwise growth), and the growth rate ratio can all be adjusted according to the characteristics of the actual dataset and the model architecture. As long as the core idea is to apply differentiated regularization strength to different body parts in order to balance stability and diversity, they all fall within the scope of protection of this application.

[0063] Example 7

[0064] This application also provides an electronic device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar as described in any of embodiments 1 to 6 above. Specifically, this electronic device constitutes the hardware implementation carrier of the technical solution of this application. The processor can be a central processing unit (CPU), graphics processing unit (GPU), neural network processing unit (NPU), digital signal processor (DSP), or field-programmable gate array (FPGA), etc., with data processing capabilities. It is responsible for calling and executing the computer program instructions stored in the memory to complete computationally intensive tasks such as timing alignment of radar point clouds, feature extraction, multiple hypothesis generation, and reliability assessment. The memory can include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device, flash memory device, or other solid-state storage device, for persistently storing the program code implementing the above method, as well as intermediate data and model parameters generated during operation. Furthermore, the electronic device may also include conventional components such as a communication bus, input / output interfaces, and network communication modules to achieve data interaction with radar sensors, auxiliary vision sensors, or cloud servers. It should be understood that the form of the electronic device is not limited. It can be an embedded computing unit integrated inside a radar module, or a standalone edge computing box, server, personal computer, or mobile terminal, as long as it has the computing power and storage resources required to execute the above methods. Through this limitation of hardware entity, the technical solution of this application can cover various types of infringing acts such as manufacturing, selling, offering for sale, and using electronic devices, providing a solid physical product protection foundation for patent rights.

[0065] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar as described in any of Embodiments 1 to 6 above. Specifically, the computer-readable storage medium is the distribution and deployment carrier of the technical solution of this application at the software level. This medium can be a tangible, non-transitory storage medium, such as a read-only memory (ROM), random access memory (RAM), a disk (e.g., floppy disk, hard disk), an optical disk (e.g., CD-ROM, DVD), a USB flash drive, a memory card, or a magnetic tape; or any other medium capable of carrying or storing program code. In some embodiments, the storage medium can also be a signal stream or carrier transmitted over a network, as long as it can be restored to executable program instructions at the receiving end. The computer program stored on this medium contains a series of instructions that, when loaded into the processor of an electronic device and executed, can drive hardware resources to process radar data according to the logical flow defined in the foregoing embodiments, thereby achieving high-precision three-dimensional posture reconstruction and reliability output. By solidifying the method and process into a program product on a storage medium, this application not only facilitates cross-platform portability and version updates of the technology, but also effectively extends the scope of protection to the circulation of software products, ensuring comprehensive protection of the technical solution in different application scenarios.

[0066] Example 8

[0067] This application also provides a specific application scenario of the above-mentioned three-dimensional reliable human posture reconstruction method based on millimeter-wave radar in a smart home environment, particularly for the monitoring of falls and abnormal postures of elderly people living alone. It should be understood that this embodiment is only used to exemplify the implementation of the technical solution in actual business and is not intended to limit the scope of protection of this application. In other feasible implementations, this method is also applicable to any radar sensing scenario requiring high precision and reliable measurement, such as sleep breathing monitoring, gait anomaly analysis, rehabilitation training assessment, or human-computer interaction gesture recognition.

[0068] In this application scenario, millimeter-wave radar is deployed in areas frequently used by the elderly, such as living rooms, bedrooms, or bathrooms, to continuously collect radar point cloud sequences within the environment. The complete processing flow described in Examples 1 to 5 is executed in real time, including temporal alignment and compensation processing of the original point cloud to generate a target point cloud sequence with compensation markers, feature extraction and multiple hypothesis generation through a pose reconstruction model, and selection of a target pose set based on human physical constraints. Finally, statistical uncertainty and radar spatial observation conditions are fused to calculate the joint-by-joint confidence level. Unlike existing technologies that only output a single pose coordinate, this scenario outputs a three-dimensional pose estimation result and its corresponding joint-by-joint confidence level, providing crucial confidence for downstream business decisions.

[0069] Specifically, when a human body configuration matching preset fall characteristics is identified based on the 3D pose estimation results, an emergency alarm is not triggered directly. Instead, the reliability of each joint in key body parts (such as the head and center of the torso) is first verified. If the reliability of these key joints is higher than a preset safety threshold, it indicates that the current pose estimation has sufficient data support and low model uncertainty. For example, if the elderly person is within the effective detection range of the radar, the point cloud density is high, and multiple candidate poses are highly consistent, then it is determined to be a real fall event, and a high-level emergency alarm is immediately sent to the guardian or emergency center via an audible and visual alarm, SMS, or cloud platform interface. This high-reliability triggering mechanism ensures a decisive response in truly critical moments, preventing delays in rescue due to excessive conservatism.

[0070] Conversely, if a suspected fall posture is detected, but the reliability of each joint of key joints such as the head or torso is below a preset safety threshold, the event is determined to be a suspected false alarm, and the system automatically enters a secondary review mode instead of directly triggering an alarm. Reasons for low reliability of key joints may include: the target being at the edge of the radar's blind zone resulting in a low prior value for radar observability; the target being severely obscured by furniture leading to a high proportion of interpolation compensation points, thus increasing statistical uncertainty; or the presence of non-human interference sources in the environment, such as fan blades or swaying curtains, creating a false point cloud structure resembling a human body. In the secondary review mode, various differentiated handling strategies can be adopted, such as automatically extending the observation time window to accumulate more frames for secondary confirmation, triggering the activation of indoor RGB cameras for visual verification, or only pushing low-priority status prompts to the user terminal for manual review afterward. This reliability-based hierarchical response mechanism fundamentally solves the industry pain points of traditional radar fall detection solutions, which suffer from high false alarm rates, poor user experience, and even forced abandonment due to a lack of reliability metrics. It significantly improves the safety and practicality of intelligent monitoring systems in actual deployment.

[0071] Furthermore, joint-by-joint credibility can also be used in this scenario to dynamically adjust the sensitivity of the posture recognition algorithm itself. For example, when the credibility of a certain area is consistently low, the posture similarity threshold for fall detection in that area can be adaptively increased, or the alarm function can be temporarily suspended until the credibility returns to a normal level, thereby proactively avoiding risks when environmental conditions deteriorate. It should be understood that the specific values ​​of the aforementioned safety thresholds, the specific strategies of the secondary review mode, and the way credibility is combined with other business logic can all be flexibly configured according to the actual installation environment, radar performance parameters, and user preferences. As long as the core idea is to use the joint-by-joint credibility output by this application to optimize the robustness and security of business decisions, it falls within the reasonable application scope of the technical solution of this application.

[0072] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for reconstructing a three-dimensional reliable human posture based on millimeter-wave radar, characterized in that, include: A radar point cloud sequence is acquired, and a time-series alignment and compensation process is performed on the radar point cloud sequence to obtain a target point cloud sequence. The detection points in the target point cloud sequence are marked with compensation identifiers indicating the data source; The target point cloud sequence is subjected to feature extraction and multiple hypothesis generation through a pose reconstruction model to obtain multiple candidate poses; during the feature extraction process, feature weight suppression is performed on the detection points from the compensation process based on the compensation identifier. Based on human physical constraints, the multiple candidate postures are screened for consistency to obtain the target posture set. Based on the statistical uncertainty of the target attitude set and the spatial observation conditions of the target joints by the radar, the joint-by-joint confidence is determined, and the three-dimensional attitude estimation result and the joint-by-joint confidence are output.

2. The method according to claim 1, characterized in that, The step of performing time alignment and compensation processing on the radar point cloud sequence to obtain the target point cloud sequence includes: In the radar point cloud sequence, a continuous empty frame segment is located, and interpolation compensation is performed based on the adjacent valid frames of the continuous empty frame segment to generate compensation detection points. The compensation identifier of the compensation detection points is marked as the interpolation compensation type. For the actual detection points in the radar point cloud sequence, the displacement compensation amount is calculated based on the radial Doppler velocity of the actual detection points, and the displacement compensation amount is superimposed on the original spatial coordinates of the actual detection points to obtain the corrected spatial coordinates aligned to the target time. For the compensation detection point whose compensation identifier is the interpolation compensation type, its interpolation coordinates remain unchanged; The target point cloud sequence is constructed based on the real detection points and the compensated detection points.

3. The method according to claim 2, characterized in that, The step of interpolating and compensating based on adjacent valid frames of the continuous empty frame segment to generate compensation detection points includes: Determine the forward valid frames before and the backward valid frames after the continuous empty frame segment; Using the spatial coordinates and radial Doppler velocity of each detection point in the forward effective frame, and according to the uniform motion model, each detection point is forward propagated to the moment of each empty frame in the continuous empty frame segment to generate forward compensation detection points. When the number of frame intervals between the forward effective frame and the backward effective frame is greater than the preset single-frame interpolation length, backward propagation is performed using the spatial coordinates and radial Doppler velocity of each detection point in the backward effective frame. The forward compensation detection point and the backward propagation detection point are weighted and fused according to the time distance ratio of each empty frame relative to the forward effective frame and the backward effective frame to obtain the compensation detection point. The calculation of displacement compensation based on the radial Doppler velocity of the actual detection point includes: Calculate the radial unit vector from the radar origin to the actual detection point based on the spatial coordinates of the actual detection point; The product of the radial Doppler velocity at the actual detection point and the radial unit vector is taken as the three-dimensional velocity vector; Calculate the time difference between the current frame and the target time, and use the product of the three-dimensional velocity vector and the time difference as the displacement compensation amount.

4. The method according to claim 1, characterized in that, The step involves feature extraction and multiple hypothesis generation of the target point cloud sequence using a pose reconstruction model to obtain multiple candidate poses, including: In the spatial feature extraction stage, each detection point in the point cloud of each frame is mapped to a point-level feature vector, and a differentiated attention mask is constructed according to the compensation identifier, so that the attention weight of the detection points not from the compensation processing is higher than the attention weight of the detection points from the compensation processing. The point-level feature vectors are aggregated into a frame-level feature vector through mask attention pooling. In the temporal coding stage, the frame-level feature vectors of each frame are arranged in chronological order. For a frame containing detection points from the compensation process, the position coding attenuation factor is determined according to the proportion of the detection points from the compensation process to the total number of detection points in the frame. The position coding attenuation factor is superimposed on the position coding to reduce the contribution weight of the frame to the temporal context. After self-attention coding, human condition features are extracted. The multiple candidate poses are generated based on the aforementioned human condition features.

5. The method according to claim 1, characterized in that, The feature extraction of the target point cloud sequence through the pose reconstruction model includes: Multi-scale feature extraction is performed on the point cloud in the target point cloud sequence to obtain spatial features at different scales; The spatial features at different scales are fused across scales to obtain fused spatial features. Subsequent temporal coding and pose generation are performed based on the fused spatial features.

6. The method according to claim 1, characterized in that, The consistency screening of the multiple candidate poses based on human physical constraints yields a target pose set, including: For each candidate pose, calculate the relative deviation between the length of each bone segment and the preset human bone segment length reference value, and calculate the inter-frame joint velocity relative to the output pose of the previous frame. Candidate poses whose relative deviation exceeds a preset bone segment deviation threshold are marked as bone segment abnormality hypotheses, and candidate poses whose inter-frame joint velocities exceed a preset velocity threshold are marked as velocity abnormality hypotheses; The target pose set is obtained by removing candidate poses that are simultaneously labeled as the bone segment abnormality hypothesis and the velocity abnormality hypothesis from the plurality of candidate poses; The determination of joint-by-joint reliability based on the statistical uncertainty of the target attitude set and the spatial observation conditions of the target joints by the radar includes: Calculate the joint-by-joint coordinate variance of multiple candidate postures in the target posture set, and use it as the statistical uncertainty; Based on the spatial position of each joint in the three-dimensional pose estimation result, the spatial distance between each detection point and each joint in the target point cloud sequence is calculated, and the radar observability prior value of each joint is estimated according to the spatial distance. The statistical uncertainty is weighted and fused with the prior value of radar observability to obtain the joint-by-joint confidence.

7. The method according to claim 6, characterized in that, The step of removing candidate poses from the plurality of candidate poses that are simultaneously labeled with the bone segment abnormality hypothesis and the velocity abnormality hypothesis to obtain the target pose set further includes: If the candidate pose set after elimination is not empty, then the candidate pose set after elimination is taken as the target pose set, and the joint-by-joint mean of the candidate poses in the target pose set is taken as the three-dimensional pose estimation result. If the candidate pose set after elimination is empty, then the multiple candidate poses are taken as the target pose set, and the joint-by-joint mean of the multiple candidate poses is taken as the three-dimensional pose estimation result. The step of weightedly fusing the statistical uncertainty with the prior value of radar observability to obtain the joint-by-joint confidence includes: When the radar observability prior value is lower than the preset observability threshold, the joint-by-joint confidence is suppressed to the preset confidence lower limit. When the radar observability prior value is not lower than the preset observability threshold, the statistical uncertainty and the radar observability prior value are weighted and summed according to a preset weight ratio to obtain the joint-by-joint confidence.

8. The method according to claim 1, characterized in that, The pose reconstruction model is trained in the following way: Simultaneously acquire radar point cloud sample sequences and image sample sequences from auxiliary vision sensors, and use a pre-trained 3D human body reconstruction model to generate initial teacher labels based on the image sample sequences; Based on the depth effective range verification and the inter-frame bone segment length consistency deviation, a joint-by-joint soft confidence weight is calculated for each joint in the initial teacher label to obtain a quality-weighted teacher label. The pose reconstruction model is trained using a time-aligned radar point cloud sample sequence as input and the quality-weighted teacher labels as the supervision target; during training, the position loss is differentially weighted according to the joint-by-joint soft confidence weights. In the multi-hypothesis generation module of the posture reconstruction model, a grouped annealing strategy is used to regularize the KL divergence loss: the human joints are divided into trunk joint group and limb distal joint group. When calculating the reconstruction loss, the trunk joint group is given a first annealing weight and the limb distal joint group is given a second annealing weight. The growth rate of the second annealing weight is slower than that of the first annealing weight.

9. An electronic device, characterized in that, include: processor; Memory, used to store computer programs; The processor is used to implement, when executing the computer program, a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a three-dimensional reliable human posture reconstruction method based on millimeter-wave radar as described in any one of claims 1 to 8.