Three-dimensional human body posture estimation method based on motion prior and diffusion model

Abstract

The invention provides a three-dimensional human body posture estimation method based on motion prior and a diffusion model. A three-dimensional human body posture estimation task is divided into a key frame denoising branch and a displacement denoising branch which are parallel; in a key frame branch, parallel injection is carried out on a two-dimensional joint point of a center frame and front and back displacement of the center frame, and then a high-precision key frame is restored by adopting space and time attention in sequence; in a displacement branch, a denoising key frame is used as a condition to be injected into a time attention layer to learn cross-frame motion, and then joint displacement is refined through space attention; independent loss functions are respectively designed for a key frame estimation network and a displacement estimation network, so that special constraints on spatial reconstruction and time coherence are realized. According to the method, the reconstruction precision of each frame of skeleton structure is ensured, and the smooth coherence of the action sequence is remarkably improved; the core problems of depth ambiguity, self-shielding, sequence jitter and the like are effectively solved, and the method has good engineering deployment value and popularization prospect.

Description

Technical Field

[0001] This invention relates to the fields of diffusion models and 3D human pose estimation, specifically to a 3D human pose estimation method based on motion priors and diffusion models. Background Technology

[0002] In recent years, generative methods based on diffusion models have gradually become an emerging research direction in the field of 3D human pose estimation (3D HPE) due to their strong probability distribution fitting ability. However, these methods still face challenges in terms of spatial localization and temporal consistency due to depth ambiguity and self-occlusion problems in monocular videos.

[0003] To address depth ambiguity, many studies have introduced prior constraints on the human skeleton or camera parameters to increase constraint information; to address temporal coherence, some researchers have used graph neural networks, recurrent neural networks, or frame-level convolutions to process motion patterns between adjacent frames. However, these methods largely rely on the network structure implicitly capturing spatial-temporal characteristics during training, making it difficult to simultaneously ensure intra-frame anatomical plausibility and cross-frame motion smoothness. Recently, diffusion models, due to their powerful random generation capabilities, have been applied to 3D human pose estimation. Some methods simultaneously add noise to the entire 3D joint sequence and introduce a spatial-temporal Transformer during denoising to achieve implicit modeling of the human skeletal structure. These methods excel in skeletal structure recovery, but because the denoiser struggles to explicitly learn the smoothness of motion trajectories, jitter and discontinuities still exist during periods of rapid motion change or static transitions.

[0004] Based on the above shortcomings, there is an urgent need for a framework that can separate spatial priors and motion priors and perform explicit denoising learning on each, so as to improve the spatial accuracy and temporal consistency of 3D pose without increasing additional inference overhead. To this end, this invention proposes a motion prior learning method based on a diffusion model, which divides the method into two branches: keyframe estimation and relative displacement estimation. By jointly modeling interval-level keyframe anchor points and cross-frame joint displacements, it achieves high-precision skeleton reconstruction and smooth transitions over long time sequences, effectively solving the jitter problem of existing diffusion methods in dynamic scenes. Summary of the Invention

[0005] Purpose of the invention: This invention proposes a three-dimensional human pose estimation method based on motion priors and a diffusion model. By explicitly learning spatial priors and motion priors through keyframe estimation branches and displacement estimation branches respectively, spatial accuracy and temporal consistency are improved simultaneously without increasing additional inference complexity.

[0006] Technical solution: The present invention provides a three-dimensional human pose estimation method based on motion priors and a diffusion model, which specifically includes the following steps:

[0007] (1) Extract key frames at fixed frame intervals from the original three-dimensional pose sequence of the input video, and calculate the relative displacement of the three-dimensional joints between adjacent frames to obtain the key frame sequence and displacement sequence respectively.

[0008] (2) Construct a dual-branch diffusion denoising model, including a keyframe denoising network and a displacement denoising network;

[0009] (3) Based on the forward process of the dual-branch diffusion denoising model, noise is added to the key frame sequence and displacement sequence with Gaussian noise coefficients in multiple time steps to generate noisy key frames and noisy displacements. At the same time, the two-dimensional key point sequence and the current diffusion step number are used as conditional inputs.

[0010] (4) Input the noisy keyframes, two-dimensional keypoint conditions and the current diffusion step into the keyframe denoising network, and perform relative displacement injection, spatial attention self-attention and inter-frame temporal attention self-attention in sequence to restore the noiseless keyframe sequence.

[0011] (5) Noisy displacement, denoised keyframe results and corresponding two-dimensional keypoint conditions input displacement denoising network, first inject spatial anchors in the temporal attention layer to explicitly learn cross-frame motion priors, then fuse joint constraints in the spatial attention layer to output a noiseless relative displacement sequence.

[0012] (6) During the training phase, independent loss functions are designed for the keyframe estimation network and the displacement estimation network respectively, so as to achieve specific constraints on spatial reconstruction and temporal coherence.

[0013] (7) During the inference phase, the keyframe denoising network and the displacement denoising network are called in turn to obtain noiseless keyframes and noiseless displacements. In each keyframe interval, the accumulated relative displacement is added to the noiseless keyframe at the starting point of the interval to restore the complete three-dimensional pose sequence frame by frame.

[0014] Furthermore, the implementation process of step (1) is as follows:

[0015] Keyframe construction uses a fixed frame interval n c The equidistant uniform sampling method, that is, from the original three-dimensional pose sequence The index M(j) is calculated as 1 + (j-1)n. c Select the j-th keyframe k0:

[0016]

[0017] The relative joint displacement sequence Δ0 between two adjacent frames:

[0018]

[0019] Where N represents the number of frames in the input two-dimensional sequence.

[0020] Further, the keyframe denoiser in step (2) includes:

[0021] The interval relative displacement injection module injects the absolute coordinates of the two-dimensional joint points in the center frame with their preceding coordinates. Frame and after The two-dimensional joint displacements of the frame relative to the central frame are embedded in parallel to provide both spatial reference and short-term motion information.

[0022] The spatial attention Transformer layer performs spatial self-attention operations on the 3D features of all joints within each keyframe to model the spatial structural relationships within the skeleton.

[0023] The temporal attention Transformer layer performs inter-frame self-attention operations on the spatial feature sequences of all keyframes to capture long-distance temporal dependencies and restore global temporal coherence.

[0024] Further, the displacement denoising network in step (2) includes:

[0025] The temporal attention Transformer layer performs self-attention operations on the noise displacement of each joint in the temporal dimension to learn cross-frame motion priors;

[0026] The spatial attention Transformer layer performs self-attention operations on the temporal features of each joint in the spatial dimension to fuse the mutual constraints between joints and refine the displacement prediction results.

[0027] Furthermore, the implementation process of step (3) is as follows:

[0028] Noise addition follows the cosine noise scheduling {β} of DDPM t At time step t, the following steps are performed on the keyframe sequence k0 and the displacement sequence Δ0 respectively:

[0029]

[0030] Where, k t To generate a noisy keyframe sequence, Δ t It is a noisy displacement sequence. α t :=1-β t , And the complete video's two-dimensional keypoint sequence Together with the current noise step number t, they serve as inputs for denoising conditions. The noise is standard Gaussian noise, and J represents the number of key points per frame.

[0031] Furthermore, the implementation process of step (4) is as follows:

[0032] Before and after each keyframe interval 2D keypoint displacement of the frame relative to the center frame:

[0033]

[0034] The absolute two-dimensional coordinates x of the center frame (M(j)) Parallel mapping to high-dimensional features is followed by concatenation; a spatial attention layer performs multi-head self-attention on the 3D features of J joints in each frame to model the skeletal topology; a temporal attention layer performs inter-frame self-attention on the spatial feature sequences of all keyframes to capture long-range temporal dependencies across keyframes, ultimately outputting denoised keyframes.

[0035] Furthermore, the loss function L of the keyframe estimation network described in step (6) Keyframe This includes keyframe location reconstruction errors and bidirectional consistency loss.

[0036] The keyframe position reconstruction error is:

[0037]

[0038] The bidirectional consistency loss is:

[0039]

[0040] final:

[0041] L Keyframe =L K +λ cons L cons

[0042] Where, λ cons To balance the hyperparameters.

[0043] Furthermore, the loss function L of the displacement estimation network described in step (6) displacement This includes displacement reconstruction error and cumulative displacement recovery error:

[0044] The displacement reconstruction error is:

[0045]

[0046] The cumulative displacement recovery error is:

[0047]

[0048] The final displacement branch loss function is defined as:

[0049] L displacement =L P +λ D L D

[0050] Where, λ D To balance the hyperparameters.

[0051] Furthermore, the implementation process of step (7) is as follows:

[0052] The keyframe estimation and displacement estimation described above are performed sequentially. The estimated keyframe k0 is then accumulated and added to the relative displacement Δ0 in the time domain according to the following formula to reconstruct the complete 3D pose sequence frame by frame.

[0053]

[0054] in, M(j)≤i <M(j+1),1≤S<n c M(j) is the y0 index corresponding to the j-th keyframe.

[0055] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a dual-branch diffusion denoising framework, this invention explicitly learns the spatial and motion priors in 3D human pose estimation. The keyframe estimation branch focuses on high-precision restoration of the skeletal structure in each frame, while the displacement estimation branch explicitly captures cross-frame motion priors, achieving natural smoothness and high spatial accuracy of motion trajectories during drastic changes in movement and transitions to stillness. The organic combination of relative displacement injection and the space-time attention layer enables the network to fully perceive local short-term motion information and global long-term temporal dependencies, thereby significantly reducing jitter and abrupt changes. At the same time, without complex post-processing, it can achieve real-time, stable, and accurate 3D human pose estimation in various application scenarios such as video surveillance, human-computer interaction, and virtual reality, balancing accuracy and efficiency. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0057] Figure 2 This is a schematic diagram of the dual-branch diffusion denoising model structure proposed in this invention;

[0058] Figure 3 This is a diagram comparing single-frame results of different methods based on diffusion models;

[0059] Figure 4 A schematic diagram comparing dynamic continuous frame results of different methods based on diffusion models;

[0060] Figure 5 A schematic diagram comparing static continuous frame results of different methods based on diffusion models; Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings.

[0062] like Figure 1 As shown, this invention proposes a three-dimensional human pose estimation method based on motion priors and a diffusion model, comprising the following steps:

[0063] Step 1: Extract keyframes from the original 3D pose sequence of the input video at fixed frame intervals, and calculate the relative displacement of the 3D joints between adjacent frames to obtain the keyframe sequence and displacement sequence, which serve as the respective denoising targets of the dual-branch network.

[0064] Keyframe construction uses a fixed frame interval n c The equidistant uniform sampling method, that is, from the original three-dimensional pose sequence The index M(j) is calculated as 1 + (j-1)n. c Select the j-th keyframe

[0065]

[0066] Relative joint displacement sequence between two adjacent frames:

[0067]

[0068] Two-dimensional keypoint sequences are not sampled and are used as complete conditional inputs. This division ensures that keyframes provide anchor points for spatial reconstruction, and displacement sequences provide a reference for cross-frame motion; the two are complementary and together cover the entire video temporal domain.

[0069] Step 2: Construct as follows Figure 2 The dual-branch diffusion denoising model shown includes a keyframe denoising network and a displacement denoising network.

[0070] The keyframe denoising network consists of three parts: an interval relative displacement injection module, which injects the absolute coordinates of the two-dimensional keypoints in the center frame into its preceding values; and an interval relative displacement injection module, which injects the absolute coordinates of the two-dimensional keypoints in the center frame into its preceding values. Frame and after The 2D joint displacements of the frames relative to the central frame are embedded in parallel to provide both spatial reference and short-term motion information. The spatial attention Transformer layer performs spatial self-attention operations on the 3D features of all joints within each keyframe to model the spatial structural relationships within the skeleton. The temporal attention Transformer layer performs inter-frame self-attention operations on the spatial feature sequences of all keyframes to capture long-distance temporal dependencies and restore global temporal coherence.

[0071] The displacement estimation branch consists of the following two parts in sequence: a temporal attention Transformer layer, which performs self-attention operations on the noise displacement of each joint in the temporal dimension to learn cross-frame motion priors; and a spatial attention Transformer layer, which performs self-attention operations on the temporal features of each joint in the spatial dimension to fuse the mutual constraints between joints and refine the displacement prediction results.

[0072] At each diffusion time step, the diffusion backbone adds Gaussian noise to the keyframe sequence and relative displacement sequence according to preset noise scheduling parameters, and calculates the noisy keyframes and noisy displacements. Simultaneously, the 2D keypoint sequence from all video frames is embedded with the current diffusion step number through position encoding and concatenated with the noise data to form the input conditions for the denoising network. This module's design ensures that the 3D target, 2D conditions, and time step information are aligned within a unified feature space, providing downstream branches with comprehensive features that include both temporal background and spatial anchors, thus laying a solid foundation for subsequent keyframe and displacement recovery.

[0073] In the keyframe estimation branch, the relative displacement injection submodule first maps the absolute 2D keypoints of each keyframe and the 2D displacements of the preceding and following half-frames in parallel to a high-dimensional feature space, and then concatenates them at the joint dimension to bring local short-term motion context. This injection method not only provides spatial anchor point information but also allows the network to perceive subtle dynamic changes in adjacent frames. Based on this, the spatial attention Transformer layer learns the geometric and anatomical constraints between joints within each frame through multi-head self-attention, enabling the network to capture the intrinsic connections of the skeletal topology and enhancing node representation capabilities using a feedforward network. Next, the temporal attention Transformer layer performs inter-frame self-attention operations on the spatial feature sequences of all keyframes, effectively capturing long-distance temporal dependencies, and finally outputs denoised 3D keyframe features to reconstruct a high-precision skeleton structure.

[0074] In the displacement estimation branch, relative motion between consecutive frames is recovered. This branch first injects the noiseless keyframe results and corresponding 2D keypoint conditions into the denoiser in the temporal attention layer, performs self-attention updates on the noisy displacement features, and explicitly learns motion priors across keyframes. Conditional injection allows attention calculation to reference the skeleton structure of the current frame to obtain more reasonable motion direction and amplitude. Subsequently, the spatial attention Transformer layer performs self-attention at the joint level, fusing the spatial constraints between joints, so that the predicted relative displacement not only conforms to the learned motion pattern but also maintains the spatial relationships of human joints, outputting the final denoised displacement features.

[0075] Step 3: Based on the forward process of the diffusion model, noise is added to the keyframe sequence and displacement sequence with Gaussian noise coefficients in multiple time steps to generate noisy keyframes and noisy displacements. At the same time, the two-dimensional keypoint sequence and the current diffusion step number are used as conditional inputs.

[0076] Noise addition follows the cosine noise scheduling {β} of DDPM t At time step t, the following steps are performed on the keyframe sequence and the displacement sequence respectively:

[0077]

[0078] in, α t :=1-β t , And the complete video's two-dimensional keypoint sequence Together with the current noise step number t, they serve as inputs for denoising conditions.

[0079] Step 4: Input the noisy keyframes, 2D keypoint conditions and the current diffusion step number into the keyframe denoising network. The network sequentially performs relative displacement injection, spatial attention self-attention and inter-frame temporal attention self-attention to restore the noise-free keyframe sequence.

[0080] The relative displacement injection module will inject the values ​​before and after each keyframe interval. 2D keypoint displacement of the frame relative to the center frame:

[0081]

[0082] The absolute two-dimensional coordinates x of the center frame (M(j)) Parallel mapping to high-dimensional features is followed by concatenation; a spatial attention layer performs multi-head self-attention on the 3D features of J joints in each frame to model the skeletal topology; a temporal attention layer performs inter-frame self-attention on the spatial feature sequences of all keyframes to capture long-range temporal dependencies across keyframes, ultimately outputting denoised keyframes.

[0083] Step 5: Input the noisy displacement, denoised keyframe results and corresponding 2D keypoint conditions into the displacement denoising network. This network first injects spatial anchors into the temporal attention layer to explicitly learn cross-frame motion priors, and then fuses inter-joint constraints in the spatial attention layer to output a noiseless relative displacement sequence.

[0084] The displacement estimation branch first injects the noiseless keyframes and their corresponding 2D keypoint conditional vectors into the temporal attention layer, and then performs the noisy relative displacement Δ. t Temporal attention operations are performed to explicitly learn cross-frame motion priors; subsequently, a spatial attention layer performs joint-level self-attention on temporal features to fuse skeletal structure constraints and refine predicted displacement.

[0085] Step 6: During the training phase, independent loss functions are designed for the keyframe estimation branch and the displacement estimation branch respectively to achieve specific constraints on spatial reconstruction and temporal coherence. During the inference phase, the keyframe and displacement denoising networks are called successively to obtain noiseless keyframes and noiseless displacements. In each keyframe interval, the accumulated relative displacement is added to the noiseless keyframe at the starting point of the interval to recover the complete 3D pose sequence frame by frame.

[0086] The training objective for the keyframe estimation branch consists of two parts: keyframe reconstruction error and bidirectional consistency error, and its specific form is as follows:

[0087]

[0088] L Kyframe =L K +λ cons L cons

[0089] Where, λ cons To balance the hyperparameters.

[0090] The training objective for the displacement estimation branch is the displacement reconstruction error L. D and cumulative attitude recovery error L P It consists of two parts, and its specific form is as follows:

[0091]

[0092] L displacement =L P +λ D L D

[0093] Where, λ D To balance the hyperparameters.

[0094] During the inference phase, 3D pose reconstruction employs a segmented accumulation strategy, combining denoised keyframes and relative displacements to recover the 3D joint trajectories of the entire sequence. Within each adjacent keyframe interval, the predicted relative displacements are sequentially accumulated, starting from the keyframe, to generate the 3D pose of each frame. This method can ensure the accuracy of keyframes while smoothly connecting continuous movements frame by frame, avoiding error accumulation across multiple intervals, and maintaining smooth transitions even during static or slow-motion phases.

[0095] For each keyframe interval (M(j), M(j+1)), press:

[0096]

[0097] in, M(j)≤i <M(j+1),1≤S<n c M(j) is the y0 index corresponding to the j-th keyframe.

[0098] This invention decomposes the 3D human pose estimation task into parallel keyframe denoising and displacement denoising branches, achieving explicit decoupling of spatial and motion priors. In the keyframe branch, the 2D joints of the center frame and their preceding and following displacements are injected in parallel, and then spatial and temporal attention is sequentially applied to reconstruct high-precision keyframes. In the displacement branch, denoised keyframes are conditionally injected into the temporal attention layer to learn cross-frame motion, and then joint displacements are refined through spatial attention. Finally, the complete 3D pose sequence is recovered through the segmented accumulation of keyframe anchors and relative displacements. This method ensures the reconstruction accuracy of the skeletal structure in each frame while significantly improving the smoothness and coherence of the action sequence.

[0099] To verify this invention, it was implemented on a GeForce RTX4090 GPU based on the PyTorch framework, using MixSTE as the backbone of the denoising unit. The channel size of the keyframe denoising unit was set to d. k =512, the channel size of the displacement noise reducer is set to d Δ =128. Using the AdamW optimizer, the momentum parameters were set to β1 = 0.9 and β2 = 0.999, and the weight decay was set to 0.1; the initial learning rate was 6 × 10⁻⁶. -5 The time decays by a factor of 0.993 after each training round; the maximum time step T is 1000; during the training phase, the keyframe interval length is set to n. c =3, displacement interval length is set to n Δ =9; hyperparameter is λ cons =1 and λ D =20.

[0100] This invention is compared with two state-of-the-art 3D human pose estimation models, D3DP (Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation) and FinePOSE (Fine-Grained Prompt-Driven 3D Human Pose Estimation via DiffusionModels). Figure 3 The comparison of single-frame results demonstrates that the present invention emphasizes temporal coherence by learning motion priors, which further guides more accurate spatial estimation and can alleviate the challenges of depth blur and self-occlusion. Figure 4The results of dynamic continuous frames are shown, and D3DP and FinePose produce disjointed motion trajectories for actions involving large movements of the hand and foot joints. Figure 5 The results of static continuous frames are shown for comparison. Both D3DP and FinePose suffer from severe jitter issues. In contrast, the reconstruction results of this invention not only generate more natural and coherent motion trajectories, but also effectively capture the semantic features of static actions.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional human pose estimation method based on motion priors and a diffusion model, characterized in that, Includes the following steps: (1) Extract key frames at fixed frame intervals from the original three-dimensional pose sequence of the input video, and calculate the relative displacement of the three-dimensional joints between adjacent frames to obtain the key frame sequence and displacement sequence respectively. (2) Construct a dual-branch diffusion denoising model, including a keyframe denoising network and a displacement denoising network; (3) Based on the forward process of the dual-branch diffusion denoising model, noise is added to the key frame sequence and displacement sequence with Gaussian noise coefficients in multiple time steps to generate noisy key frames and noisy displacements. At the same time, the two-dimensional key point sequence and the current diffusion step number are used as conditional inputs. (4) Input the noisy keyframes, two-dimensional keypoint conditions and the current diffusion step into the keyframe denoising network, and sequentially perform relative displacement injection, spatial attention self-attention and inter-frame temporal attention self-attention to restore the noiseless keyframe sequence. (5) Noisy displacement, denoised keyframe results and corresponding two-dimensional keypoint conditions input displacement denoising network, first inject spatial anchors in the temporal attention layer to explicitly learn cross-frame motion priors, then fuse joint constraints in the spatial attention layer to output a noiseless relative displacement sequence. (6) During the training phase, independent loss functions are designed for the keyframe estimation network and the displacement estimation network respectively, so as to achieve specific constraints on spatial reconstruction and temporal coherence. (7) During the inference phase, the keyframe denoising network and the displacement denoising network are called in turn to obtain noiseless keyframes and noiseless displacements. In each keyframe interval, the accumulated relative displacement is added to the noiseless keyframe at the starting point of the interval to restore the complete three-dimensional pose sequence frame by frame.

2. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The implementation process of step (1) is as follows: Keyframe construction uses a fixed frame interval n c The equidistant uniform sampling method, that is, from the original three-dimensional pose sequence The index M(j) is calculated as 1 + (j-1)n. c Select the j-th keyframe k0: The relative joint displacement sequence Δ0 between two adjacent frames: Where N represents the number of frames in the input two-dimensional sequence.

3. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The keyframe denoiser in step (2) includes: The interval relative displacement injection module injects the absolute coordinates of the two-dimensional joint points in the center frame with their preceding coordinates. Frame and after The two-dimensional joint displacements of the frame relative to the central frame are embedded in parallel to provide both spatial reference and short-term motion information. The spatial attention Transformer layer performs spatial self-attention operations on the 3D features of all joints within each keyframe to model the spatial structural relationships within the skeleton. The temporal attention Transformer layer performs inter-frame self-attention operations on the spatial feature sequences of all keyframes to capture long-distance temporal dependencies and restore global temporal coherence.

4. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The displacement denoising network in step (2) includes: The temporal attention Transformer layer performs self-attention operations on the noise displacement of each joint in the temporal dimension to learn cross-frame motion priors; The spatial attention Transformer layer performs self-attention operations on the temporal features of each joint in the spatial dimension to fuse the mutual constraints between joints and refine the displacement prediction results.

5. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The implementation process of step (3) is as follows: Noise addition follows the cosine noise scheduling {β} of DDPM t At time step t, the following steps are performed on the keyframe sequence k0 and the displacement sequence Δ0 respectively: Where, k t To generate a noisy keyframe sequence, Δ t It is a noisy displacement sequence. And the complete video's two-dimensional keypoint sequence Together with the current noise step number t, they serve as inputs for denoising conditions. The noise is standard Gaussian noise, and J represents the number of key points per frame.

6. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The implementation process of step (4) is as follows: Before and after each keyframe interval 2D keypoint displacement of the frame relative to the center frame: The absolute two-dimensional coordinates x of the center frame (M(j)) Parallel mapping to high-dimensional features is followed by concatenation; a spatial attention layer performs multi-head self-attention on the 3D features of J joints in each frame to model the skeletal topology; a temporal attention layer performs inter-frame self-attention on the spatial feature sequences of all keyframes to capture long-range temporal dependencies across keyframes, ultimately outputting denoised keyframes.

7. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The loss function L of the keyframe estimation network in step (6) Keyframe This includes keyframe location reconstruction errors and bidirectional consistency loss. The keyframe position reconstruction error is: The bidirectional consistency loss is: final: L Keyframe =L K +λ cons L cons Where, λ cons To balance the hyperparameters.

8. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The loss function L of the displacement estimation network in step (6) displacement This includes displacement reconstruction error and cumulative displacement recovery error: The displacement reconstruction error is: The cumulative displacement recovery error is: The final displacement branch loss function is defined as: L displacement =L P +λ D L D Where, λ D To balance the hyperparameters.

9. The three-dimensional human pose estimation method based on motion priors and a diffusion model according to claim 1, characterized in that, The implementation process of step (7) is as follows: The keyframe estimation and displacement estimation described above are performed sequentially. The estimated keyframe k0 is then accumulated and added to the relative displacement Δ0 in the time domain according to the following formula to reconstruct the complete 3D pose sequence frame by frame. in, M(j) is the y0 index corresponding to the j-th keyframe.

Images