Laser radar human motion capture method based on bezier curve degeneration modeling
By employing Bézier curve degradation modeling and a progressive reconstruction strategy, and utilizing LiDAR point cloud sequences for 3D human motion capture, this approach solves the problem of deployment difficulties in open scenes encountered by existing methods, achieving high-precision and highly coherent 3D human motion reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2026-01-15
- Publication Date
- 2026-06-09
AI Technical Summary
Existing 3D human motion capture methods are difficult to deploy in open scenes, limited by expensive equipment, sensitive to changes in lighting and lacking robustness, and lack reliable absolute depth information, resulting in pose jitter and prediction failure.
A LiDAR human motion capture method based on Bézier curve degradation modeling is adopted. By constructing a multi-level motion representation, a progressive reconstruction from coarse to fine is performed using LiDAR point cloud sequences. Combined with multi-timescale motion transformation and inverse kinematics solution, high-precision and highly coherent 3D human motion is reconstructed.
It significantly improves the robustness and accuracy of motion capture in complex open scenes, solves the problem of posture jitter caused by occlusion and noise interference, and achieves efficient 3D human motion reconstruction without relying on wearable devices and controlled lighting environments.
Smart Images
Figure CN122172212A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision technology, and in particular relates to a lidar human motion capture method based on Bézier curve degradation modeling. Background Technology
[0002] 3D human motion capture is a fundamental and valuable research task in the field of computer vision. Its core objective is to reconstruct a standardized 3D representation of the human body that changes over time from sensor data. This technology plays a crucial role in human-centered downstream applications such as autonomous driving, robotics, and augmented reality. For example, in autonomous driving systems, accurately perceiving the movement intentions of pedestrians is essential for path planning and obstacle avoidance; in robotics, understanding human actions is fundamental to achieving human-machine collaboration.
[0003] Currently, mainstream motion capture methods can be divided into several categories: The first category is the traditional marker-based or wearable device method, which obtains high-precision attitude information by attaching markers to the human body or wearing inertial measurement unit (IMU) sensors. The second category is vision-based methods, which use RGB or RGB-D cameras as input. With the development of deep learning, this type of method has achieved great success.
[0004] Despite significant advancements in the aforementioned methods, they fail to address the following issues: first, they require expensive specialized equipment and complex calibration processes, and rely on wearable hardware, making large-scale deployment in everyday or open environments difficult; second, they are limited by indoor environments and sensitive to changes in lighting; and third, they lack reliable absolute depth information, resulting in insufficient robustness in complex outdoor scenes. Therefore, a novel framework based on Bézier curve degradation modeling is needed for 3D human motion capture. Summary of the Invention
[0005] The purpose of this invention is to address the problems mentioned in the background technology by proposing a LiDAR human motion capture method based on Bézier curve degradation modeling. By constructing a multi-level motion representation with Bézier curves as the core parameterization tool and adopting a progressive reconstruction strategy from coarse to fine, the method acquires three-dimensional human motion information from a continuous multi-frame LiDAR point cloud sequence. This solves the problems of pose jitter and prediction failure faced by existing technologies when dealing with LiDAR data sparsity, occlusion, and noise interference, and ultimately achieves high-precision and highly coherent three-dimensional human motion reconstruction.
[0006] To achieve the objective of this invention, a laser radar method for human motion capture based on Bézier curve degradation modeling is disclosed, comprising the following steps:
[0007] Step 1: Obtain a continuous multi-frame LiDAR point cloud sequence;
[0008] Step 2: Construct a trajectory-aware Bezier motion degradation module to fit the original joint trajectory and gradually reduce the control points through a trajectory preservation strategy to generate a multi-level motion representation from coarse to fine.
[0009] Step 3: Design a progressive motion reconstruction module. Utilize the multi-timescale motion transformer (TMT) to predict Bezier motion curves at multiple time scales based on point cloud features. Then, use the multi-level motion aggregator (MMA) to adaptively fuse the multi-scale curves to reconstruct a detailed and temporally coherent three-dimensional human pose sequence.
[0010] Step 4: Use an inverse kinematics solver to convert the reconstructed joint positions into pose parameters of a standard human body model, and output the final three-dimensional human motion.
[0011] Furthermore, the lidar point cloud sequence in step 1 is captured within a continuous time window. Frame 3D point cloud data, each frame of point cloud is represented as ,in The number of points in each frame of the point cloud. .
[0012] Furthermore, the specific process of constructing the trajectory-aware Bézier motion degradation module in step 2 is as follows: First, constrain each joint with a cubic Bézier curve to ensure that the curve satisfies C¹ continuity at the control points; then, by setting the initial acceleration to zero, solve all control points using the Thomas algorithm to obtain the finest Bézier chain; finally, by selecting new anchor points and adjusting the control point length, downsample the Bézier chain to form degradation versions with different time resolutions, thereby constituting a multi-level motion representation.
[0013] Furthermore, the trajectory preservation strategy in step 2 is as follows: when downsampling the Bezier chain, not only are the anchor points resampled, but the unit tangent vector is also extracted from the finest curve, and the position and length of the new control point are adjusted based on this to better preserve the dynamic characteristics of the original motion.
[0014] Furthermore, the multi-timescale motion transformer (TMT) in step 3 is a Transformer with an encoder architecture that treats each layer of motion representation as an independent sequence of labels and jointly models its interaction with LiDAR features to output reconstructed motion curves at each timescale.
[0015] Furthermore, in step 3, the multi-timescale motion transformer (TMT) applies a block causal mask on the self-attention layer, allowing each motion marker to focus only on all markers from the coarser level and all point feature markers, to ensure that the coarse-grained motion trend can effectively guide the refinement of the fine-grained motion.
[0016] Furthermore, the multi-level motion aggregator (MMA) in step 3 adopts a decreasing mechanism to gradually integrate motion representations from different time scales: for the coarser level representation, the predicted Bézier curve parameters are first used for upsampling to match the length of the fine level, and then the two representations are fused through the multilayer perceptron (MLP) to finally output the joint position prediction of the finest level.
[0017] Furthermore, the inverse kinematics solver in step 4 is a solver based on the graph convolutional network ST-GCN, used to convert the estimated joint positions into pose parameters of the SMPL human model.
[0018] Compared with existing technologies, the significant advancements of this invention are: 1) This invention utilizes Bézier curves to parametrically model motion, fundamentally ensuring the physical rationality and smoothness of the reconstructed motion, and effectively alleviating the jitter or failure problems caused by occlusion; 2) The coarse-to-fine reconstruction framework proposed in this invention can fully utilize coarse-grained motion trends to compensate for observation gaps, significantly improving the robustness of the model in complex open scenes; 3) This invention does not rely on wearable devices or controlled lighting environments, but only requires monocular LiDAR input, and can be deployed and applied in real complex scenarios such as autonomous driving, robotics, and intelligent monitoring.
[0019] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description
[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0021] Figure 1 This is a schematic diagram of the process of the present invention;
[0022] Figure 2 This is a schematic diagram of the trajectory-aware Bezier motion degradation module in this invention;
[0023] Figure 3 This is a schematic diagram of the progressive motion reconstruction module in this invention;
[0024] Figure 4 This is a schematic diagram of the attention mechanism of the multi-timescale motion transformer (TMT) in this invention;
[0025] Figure 5 This is a schematic diagram of sequential visual comparison in a severely occluded scene according to the present invention;
[0026] Figure 6 This is a schematic diagram illustrating the static visual comparison of the present invention on different datasets;
[0027] Figure 7 This is a schematic diagram of the robustness test results of the present invention under different frame loss ratios;
[0028] Figure 8 This is a visualization diagram of the multi-level Bézier curves predicted by the present invention. Detailed Implementation
[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] A lidar-based human motion capture method based on Bézier curve degradation modeling includes the following steps:
[0031] Step 1: Obtain a continuous multi-frame LiDAR point cloud sequence;
[0032] Step 2: Construct a trajectory-aware Bezier motion degradation module to fit the original joint trajectory and gradually reduce the control points through a trajectory preservation strategy to generate a multi-level motion representation from coarse to fine.
[0033] Step 3: Design a progressive motion reconstruction module. Use a multi-timescale motion transformer (TMT) to predict Bezier motion curves at multiple time scales based on point cloud features. Then, use a multi-level motion aggregator (MMA) to adaptively fuse the multi-scale curves to reconstruct a detailed and temporally coherent three-dimensional human pose sequence.
[0034] Step 4: Use an inverse kinematics solver to convert the reconstructed joint positions into pose parameters of a standard human body model, and output the final three-dimensional human motion.
[0035] Specifically, the lidar point cloud sequence in step 1 is captured within a continuous time window. Frame 3D point cloud data, each frame of point cloud is represented as ,in The number of points in each frame of the point cloud. .
[0036] Specifically, the process of constructing the trajectory-aware Bézier motion degradation module in step 2 is as follows: First, constrain each joint with a cubic Bézier curve to ensure that the curve satisfies the following conditions at the control points. Continuity (i.e., continuity of position and first derivative); then, by setting the initial acceleration to zero and using the Thomas algorithm to solve for all forward and backward control points, the finest Bezier chain is obtained; finally, by selecting new anchor points and adjusting the control point lengths, the original Bezier chain is downsampled to form degenerate versions with different time resolutions.
[0037] Specifically, the trajectory preservation strategy in step 2 is as follows: when downsampling the Bezier chain, not only are the anchor points resampled, but the unit tangent vector is also extracted from the finest curve, and the position and length of the new control point are adjusted based on this.
[0038] Specifically, the optimal control point length parameter is solved by a least-squares optimization problem, so that the degenerate Bézier curve approximates the true path of the original trajectory in the corresponding time interval as closely as possible. This reduces the number of control points while preserving the dynamic characteristics of the original motion to the greatest extent.
[0039] Specifically, the progressive motion reconstruction module in step 3 includes two core sub-modules: the multi-timescale motion transformer (TMT) and the multi-level motion aggregator (MMA).
[0040] Specifically, TMT is a Transformer with an encoder structure that treats each layer of motion representation as an independent sequence of labels and jointly models its interaction with LiDAR features to output reconstructed motion curves at various time scales.
[0041] Specifically, MMA employs a top-down fusion mechanism, upsampling the coarser-grained representation using Bezier curve parameters and fusing features with the next layer of finer representation to gradually recover a high-resolution pose sequence.
[0042] Specifically, the multi-timescale motion transformer (TMT) in step 3 applies a block-wise causal mask to the self-attention layer, allowing each motion label to focus only on all labels from the coarser level and all point feature labels. This mechanism ensures that during model inference, coarse-grained motion trends can effectively guide the prediction of fine-grained motion, preventing fine-grained predictions from deviating from the overall motion semantics.
[0043] Specifically, in step 3, the LiDAR feature extraction uses a pre-trained PointNet++ network to encode each frame of point cloud data and maps it to a unified feature dimension through a multilayer perceptron (MLP) to form an observation condition tensor. It is used for conditional reasoning in the TMT module.
[0044] Specifically, the inverse kinematics solver in step 4 is a solver based on a spatial-temporal graph convolutional network (ST-GCN), whose input is the joint position sequence output by the MMA and whose output is the pose parameters of the SMPL human model. Furthermore, the attitude parameters are reprojected into joint positions through SMPL forward kinematics calculations for end-to-end joint optimization.
[0045] Specifically, this invention employs a multi-level supervision strategy, including: applying L2 loss to the Bezier motion curves at each level of the TMT output; applying parametric regression loss to the pose parameters output by the inverse kinematics solver; applying forward kinematics loss to the error between the SMPL forward kinematics output and the ground truth joint position; and weighted summing of the three losses to form the overall training objective, ensuring that the model is optimized simultaneously at the trajectory, parametric, and geometric levels. This method is applicable to complex open scenarios such as autonomous driving and robotics, and can alleviate pose jitter or failure caused by occlusion, noise, and sparse point clouds, significantly improving the accuracy and temporal continuity of motion capture.
[0046] Example
[0047] like Figure 1 As shown, this invention proposes a LiDAR human motion capture method based on Bézier curve degradation modeling. This method aims to reconstruct high-precision, high-temporal-coherence 3D human motion using sparse point cloud sequences acquired by LiDAR sensors in open environments. Addressing the problem that existing methods are prone to pose jitter and even prediction failure under severe occlusion and point cloud noise, this invention introduces Bézier curves for parametric modeling of joint trajectories and designs a hierarchical motion representation and reconstruction mechanism from coarse to fine, thereby significantly improving the robustness and stability of the model in complex real-world scenarios. The specific steps are as follows:
[0048] Step 1: Acquire a continuous multi-frame LiDAR point cloud sequence. In this embodiment, the input data comes from a monocular LiDAR sensor installed on an autonomous vehicle or at a fixed monitoring point. Frames of 3D point clouds within a continuous time window are collected to form the input sequence:
[0049] ,
[0050] Where N represents the number of points contained in each frame of the point cloud (typically fluctuating between hundreds and thousands), and t is the time index. This point cloud data is collected in real-world mixed indoor and outdoor scenes (such as streets, squares, and indoor corridors), inevitably containing noise points, background interference, and partial human figure loss due to pedestrian occlusion or fixed obstacles, fully conforming to actual deployment conditions. To ensure subsequent processing efficiency, each frame of the point cloud can undergo preprocessing operations such as voxel downsampling or farthest-point sampling (FPS) before input to balance the point cloud density.
[0051] Step 2: Construct a trajectory-aware Bézier motion degradation module. For example... Figure 2 As shown, this module first performs cubic Bézier curve fitting on the ground-value joint trajectories in the training set (e.g., 3D joint sequence obtained from backprojection of the SMPL model). For any joint constraint... A cubic Bézier curve is required to satisfy the following conditions at each control point: Continuity (i.e., continuity of position and first derivative) is achieved, and by setting the initial acceleration to zero, the Thomas algorithm is used to analytically solve for all forward and backward control points, thus obtaining the finest Bezier chain. Subsequently, to construct a multi-level supervision signal, this module introduces a trajectory-aware degradation strategy: the original Bezier chain is downsampled at different step sizes (e.g., ... Multi-level compression is performed; in each level of compression, not only are anchor points reselected, but the unit tangent vector is also extracted from the finest curve, and the position and length of the new control points are adjusted based on this. Furthermore, the control point length parameter is solved using least-squares optimization, ensuring that the degenerate Bézier curve approximates the original trajectory as closely as possible within the corresponding time period. Finally, a set of multi-level motion representations, from coarse to fine, is generated.
[0052]
[0053] in This refers to the number of joints (e.g., 24). This represents the three-dimensional coordinates (3×3) of each anchor point and its preceding and following control points. This representation preserves the macroscopic trend of the original motion while possessing good learning-friendliness, providing structured supervision for subsequent reconstruction.
[0054] Step 3: Design a progressive motion reconstruction module. For example... Figure 3 As shown, this module first uses a pre-trained PointNet++ network to process the point cloud for each frame. Feature extraction is performed to obtain high-dimensional point features, which are then mapped to a uniform dimension using a two-layer MLP. The feature vectors are finally concatenated into the observation condition tensor. Subsequently, multi-level motion was embedded. and Together, they are input into a Time-scale Motion Transformer (TMT). The TMT employs a standard 12-layer Transformer encoder structure, with its key innovation being the application of a block-wise causal mask to its self-attention layer. For example... Figure 4 As shown, this mask ensures that the first Any motion marker in a layer can only focus on coarser layers ( to The model focuses on all labels and point cloud feature labels at each layer, preventing the model from focusing on finer layers. This design forces the model to follow a coarse-to-fine information flow, allowing coarse-grained motion trends to effectively guide the generation of fine-grained details. TMT outputs motion curve predictions at various time scales. Then, the predictions are fused by a multi-level motion aggregator (MMA): MMA uses a top-down fusion mechanism, upsampling (i.e., resampling on the curve) the predictions of coarser levels using their Bezier parameters to align their time lengths with the next level, and then fusing the two through an MLP; this process is progressive, ultimately outputting the most refined level of joint position predictions. .
[0055] Step 4: The reconstructed joint positions are converted into pose parameters of a standard human body model using an inverse kinematics solver. Since the ultimate goal of this invention is to output pose parameters compatible with standard human body models such as SMPL, an inverse kinematics (IK) solver based on a spatial-temporal graph convolutional network (ST-GCN) is introduced after obtaining the final joint positions. This solver takes the predicted joint sequence as input, learns the kinematic constraints between joints through a multi-layer STGCN module, and outputs the corresponding SMPL pose parameters for each frame. To ensure geometric consistency and achieve end-to-end training, the SMPL forward kinematics is further utilized. Reprojection as joint position This is used for subsequent loss calculations.
[0056] Step 5: Model Training and Loss Function Calculation. This invention employs a multi-level joint supervision strategy: First, the Bezier motion representations of each level of the TMT output are trained. With truth value The Frobenius norm between the two is used to calculate the motion representation loss. Secondly, the SMPL attitude parameters output by the IK solver... With truth value Calculate parametric regression loss Finally, the joint positions of the SMPL forward kinematic reprojection were determined. Joint position with truth value Calculate positive kinematic loss The total loss function is defined as follows: The weights are set to The model was trained on four NVIDIA RTX 4090 GPUs using the AdamW optimizer with an initial learning rate of 2.5 × 10⁻ 4 Training for 50 rounds, time window length The default is 32 frames per second.
[0057] This invention, through the specific implementation methods described above, constructs a complete closed loop from Bézier curve fitting and degradation to progressive reconstruction and parameter transformation. Experiments demonstrate that this method significantly outperforms existing technologies on four mainstream datasets: LiDARHuman26M, FreeMotion, NoiseMotion, and SLOPER4D. It maintains stable output even under extreme occlusion conditions with up to 50% frame loss, effectively solving the pose jitter problem caused by data instability in traditional methods. It possesses great potential for deployment and application in practical scenarios such as autonomous driving, robot navigation, and intelligent security.
[0058] Table 1. Comparison of results between the proposed method and other existing methods on four mainstream benchmark datasets.
[0059] method LiDARHuman26M FreeMotion NoiseMotion SLOPER4D MOVIN 76.2 106.7 50.4 47.7 LiDAR-HMR 79.3 86.3 52.6 49.7 LiDARCap 75.7 85.5 62.4 71.6 LIP* 76.8 62.5 48.8 60.1 NE-LiDARCap* 71.9 69.4 48.4 96.8 Our method 66.8 47.2 36.9 36.5
[0060] Table 1 compares the proposed method (BMLiCap) with other existing methods on four mainstream LiDAR human motion capture benchmark datasets. The MPJPE metric is indicated by a "↓" sign, with lower values representing better performance. It can be seen that the proposed method achieves the best ranking on all three metrics across all four datasets (highlighted in bold), demonstrating its ability to significantly improve the accuracy and temporal continuity of reconstructing 3D human motion from LiDAR data. Particularly on the most challenging FreeMotion dataset, the proposed method achieves a significant improvement of up to 14.7 in MPJPE compared to the previous best method, LiveHPS++.
[0061] Figure 5 This is a schematic diagram illustrating sequential visual comparison in a partially severely occluded scene. Figure 5 As can be seen, when the input point cloud is severely incomplete due to body movement, existing methods such as LiDARCap and LiveHPS cannot effectively capture motion, and their outputs either jitter violently or fail completely. In contrast, this invention can consistently provide stable, accurate, and coherent motion reconstruction.
[0062] Figure 6This is a qualitative comparison between the algorithm proposed in this invention and existing baseline methods under different degrees of occlusion. The figure shows multiple samples containing significant body part occlusion or unusual movements. Figure 6 As can be seen, other methods generally suffer from jitter or prediction failure when dealing with these extreme cases, while our method exhibits excellent robustness, with stable and consistent output results.
[0063] Figure 7 This diagram illustrates the robustness test results of the present invention under different frame loss ratios. The diagram shows that even when the input point cloud frames are randomly lost by up to 50%, the present invention maintains stable performance, with its MPJPE and Accel Err indices changing much less than other comparative methods.
[0064] Figure 8 This diagram illustrates the visualization results of Bézier curve fitting of motion trajectories at different scales (k=8, 16, 32) according to the present invention. The diagram shows the gradual process of the fitted curve from coarse to fine as the number of control points k increases: Scale 1 (k=8) captures the main trend of motion but loses details; Scale 2 (k=16) significantly improves the smoothness of the trajectory and the ability to restore local features; Scale 3 (k=32) can more accurately approximate the original point cloud trajectory, demonstrating the ability of this method to achieve high-fidelity motion capture through multi-scale Bézier curve degradation modeling.
[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A lidar human motion capture method based on Bézier curve degradation modeling, characterized in that, Includes the following steps: Step 1: Obtain a continuous multi-frame LiDAR point cloud sequence; Step 2: Construct a trajectory-aware Bezier motion degradation module to fit the original joint trajectory and gradually reduce the control points through a trajectory preservation strategy to generate a multi-level motion representation from coarse to fine. Step 3: Design a progressive motion reconstruction module. Utilize the multi-timescale motion transformer (TMT) to predict Bezier motion curves at multiple time scales based on point cloud features. Then, use the multi-level motion aggregator (MMA) to adaptively fuse the multi-scale curves to reconstruct a detailed and temporally coherent three-dimensional human pose sequence. Step 4: Use an inverse kinematics solver to convert the reconstructed joint positions into pose parameters of a standard human body model, and output the final three-dimensional human motion.
2. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The lidar point cloud sequence mentioned in step 1 is captured within a continuous time window. Frame 3D point cloud data, each frame of point cloud is represented as ,in The number of points in each frame of the point cloud. .
3. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The specific process of constructing the trajectory-aware Bézier motion degradation module in step 2 is as follows: First, constrain each joint with a cubic Bézier curve to ensure that the curve satisfies C¹ continuity at the control points; then, by setting the initial acceleration to zero, solve all control points using the Thomas algorithm to obtain the finest Bézier chain; finally, by selecting new anchor points and adjusting the control point length, downsample the Bézier chain to form degradation versions with different time resolutions, thereby constituting a multi-level motion representation.
4. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The trajectory preservation strategy described in step 2 is as follows: when downsampling the Bezier chain, not only are the anchor points resampled, but the unit tangent vector is also extracted from the finest curve, and the position and length of the new control point are adjusted based on this to better preserve the dynamic characteristics of the original motion.
5. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The multi-timescale motion transformer (TMT) described in step 3 is a Transformer with an encoder architecture that treats each layer of motion representation as an independent sequence of labels and jointly models its interaction with LiDAR features to output reconstructed motion curves at each timescale.
6. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The multi-timescale motion transformer (TMT) described in step 3 applies a block causal mask on the self-attention layer, allowing each motion marker to focus only on all markers from the coarser level and all point feature markers, so as to ensure that the coarse-grained motion trend can effectively guide the refinement of fine-grained motion.
7. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The multi-level motion aggregator (MMA) described in step 3 employs a decreasing mechanism to gradually integrate motion representations from different time scales: for coarser-level representations, the predicted Bézier curve parameters are first used for upsampling to match the length of the fine-level representations, and then the two representations are fused through a multilayer perceptron (MLP) to finally output the joint position prediction of the finest level.
8. The laser radar human motion capture method based on Bézier curve degradation modeling according to claim 1, characterized in that, The inverse kinematics solver mentioned in step 4 is a solver based on the graph convolutional network ST-GCN, used to convert the estimated joint positions into pose parameters of the SMPL human model.