Three-dimensional human motion prediction method and system with global navigation and local interaction

Abstract

The application discloses a kind of global navigation and local interaction three-dimensional human motion prediction method and system.The method comprises:1, construct sparse global trajectory planner.2, construct adaptive self-centered scene extractor.3, construct progressive motion decoder.The application obtains coarse-grained global path by sparse global trajectory planner, dynamically extracts motion-related local geometric information from dense point cloud by adaptive self-centered scene extractor, and gradually generates future human motion sequence by progressive motion decoder.Compared with the method of only relying on sparse scene or complete global scene modeling, the application can reduce the computing overhead and suppress irrelevant spatial noise while retaining the key obstacle boundary around human and interaction geometric details, thereby improving the accuracy of predicted trajectory, the naturalness of motion posture and the physical reasonableness of human-environment interaction.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, and particularly to a method for predicting human motion in 3D scenes that coordinates global path navigation and local fine-grained interaction. Specifically, it is a 3D human motion prediction technology based on a hierarchical framework that combines global sparse trajectory planning with adaptive human-centered local scene modeling. This method can be widely applied to intelligent systems that require understanding and predicting human motion behavior in 3D environments, such as autonomous driving, augmented reality / virtual reality, human-computer interaction, service robots, and intelligent monitoring. Background Technology

[0002] With the rapid development of virtual reality, augmented reality, intelligent robots, and autonomous driving, accurately predicting future human actions in three-dimensional environments has become a hot research topic. Human motion prediction based on 3D scene perception aims to predict human posture sequences over a future period based on observed action sequences and 3D environmental information. This technology requires the model to possess two capabilities simultaneously: navigation in a vast spatial layout and fine-grained interaction with local geometry.

[0003] Traditional human motion prediction methods largely rely on sequence modeling architectures such as recurrent neural networks (RNNs) and graph convolutional networks (GCNs), which excel at capturing temporal dependencies and skeletal correlations. However, they typically treat human motion as a scene-independent dynamic process. Therefore, these methods struggle to generate physically plausible motion in complex 3D environments.

[0004] In recent years, researchers have begun to introduce scene-aware prediction methods, integrating environmental geometric information to constrain human motion. However, existing methods still face fundamental bottlenecks in coordinating global path navigation and fine-grained local interactions: most advanced frameworks employ aggressive scene downsampling strategies to alleviate the computational explosion caused by processing dense 3D point clouds, resulting in physically unreliable motion in cluttered environments; global scene encoding contains spatial redundancy and noise, and modeling the entire scene introduces too much task-irrelevant noise, diluting the informative motion-scene correlation and introducing redundant computational overhead, hindering the model's ability to focus on the most relevant local interactions. Therefore, there is an urgent need for a 3D human motion prediction method that can simultaneously balance global navigation accuracy and fidelity of local geometric details. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a three-dimensional human motion prediction method and system that combines global navigation and local interaction.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: A three-dimensional human motion prediction method and system based on global navigation and local interaction, specifically including the following steps: Step (1) Construct a sparse global trajectory planner: Construct a global occupancy map based on sparse scene point cloud, use the global occupancy map as a global geometric prior, and combine historical human motion sequences, sparse scene features and gaze point information to predict coarse-grained global trajectories in future time steps. The historical human motion sequences and fixation point information can be obtained directly.

[0007] Step (2): Construct an adaptive egocentric scene extractor: Determine the estimated position of the human body in the future time step based on the coarse-grained global trajectory, and dynamically adjust the size of the local perception area centered on the human body based on the instantaneous velocity of the human body. Extract local sub-scenes related to the current motion from the dense global scene point cloud to obtain the egocentric local scene features corresponding to each local sub-scene. Step (3) Construct a progressive motion decoder: Align the filling sequence with the coarse-grained global trajectory, and fuse motion features and local scene features within a sliding window. Use an autoregressive approach to gradually predict future human motion sequences, and stitch the future human motion sequences with historical human motion sequences to finally generate a 3D human motion that conforms to global path navigation constraints and local scene geometric constraints.

[0008] Further, in step (1), the sparse scene point cloud is discretized into a two-dimensional grid, and a representative height is estimated for each grid block to characterize the geometric structure of the grid block in the vertical direction. When there are scene points in the grid block, the maximum height of the scene point in the grid block is taken as the representative height of the grid block. When the grid block is empty, the space is smoothly filled according to the representative heights of its four neighboring grid blocks. Based on the comparison between the representative height and the preset height threshold, the occupancy status of each grid block is determined to obtain the global occupancy map.

[0009] When there is a scene point in the grid block (i,j), that is : (Formula 1) When the grid block (i,j) is empty, i.e. Utilizing neighborhood smoothing: (Formula 2) in, This represents the set of points that fall within grid block (i,j); Represents the set of surrounding grid blocks of grid block (i,j). Represents the neighborhood grid block The representative height; Global Occupancy Map Derived Based on Representative Height : (Formula 3) in, As an indicator function, when the representative height Exceeding the threshold At that time, the global occupation map will be... The occupied state is set to 1.

[0010] Furthermore, step (1) also includes: The constructed global occupancy graph The input is then fed into a multilayer perceptron for feature extraction, extracting the dominant features. ; The input historical human motion sequence is encoded into a raw motion embedding and then combined with the occupation feature. Fusion is performed to obtain motion characteristics; The sparse scene point cloud is encoded into global sparse scene features, and attention computation is performed on the global sparse scene features based on the gaze point information to obtain the target region representation; Multimodal fusion of motion features, global sparse scene features, and target region representation is performed to obtain interactive scene features.

[0011] By inputting motion features and interaction scene features into a Transformer-based trajectory decoder, global translation and global orientation can be predicted within a future time period.

[0012] Furthermore, the adaptive egocentric scene extractor in step (2) employs a velocity-driven local perception region scaling strategy. The local perception region is a cylindrical region centered on the estimated position of the human body, and its radius is dynamically determined based on the instantaneous velocity of the human body.

[0013]

[0014] in, This represents the local sensing radius at the k-th time step. Indicates the basic sensing radius. Indicates the radius expansion coefficient. This represents the instantaneous velocity of the human body at the k-th time step. Extracted subset. Points satisfying the following conditions Composition: The horizontal distance from the center of the cylindrical region does not exceed And the vertical coordinate is subject to a height threshold. limit, . , These represent the x and y coordinates at the current k-th time step, respectively. This represents a dense global point cloud.

[0015] Furthermore, step (2) adopts a dual-branch coding scheme, specifically as follows: From the extracted subset Uniform sampling Each point is used to construct a standardized local scene. The standardized local scene is feature-encoded using a dual-branch coding scheme. The geometric prior coding branch enhances point features using height-constrained coding and human proximity coding. The feature extraction branch extracts local geometric features using a scene encoder. The features output from both branches are then aggregated to obtain the time step. Corresponding self-centered scenario representation .

[0016] The geometric prior coding branch constructs geometric prior features based on the spatial geometric relationships of local scene points. Among them, the height constraint coding is used to describe the vertical spatial relationship of scene points relative to the ground, and the human proximity coding is used to characterize the distance relationship between scene points and the center of the human body, thereby enhancing the ability of local scene features to express information related to human motion interaction.

[0017] The feature extraction branch extracts point-level geometric features of the normalized local scene through the scene encoder (PointNet++).

[0018] Furthermore, the progressive regression described in step (3) specifically refers to: 3-1. The progressive motion decoder copies the last frame of the historical human motion sequence, thus transforming the historical human motion sequence... Time fill to To obtain the initial motion sequence Then fill the sequence. By aligning with the coarse-grained global trajectory and updating the global translation and global orientation, a motion sequence aligned with the coarse-grained global trajectory is obtained. .

[0019] 3-2. To avoid error accumulation and motion drift when predicting long motion sequences at once, and to enable the model to gradually combine the motion results and scene constraint information from the previous stage during the prediction process, a variable time step is introduced. By controlling the length of each new prediction sequence, the future motion prediction process is transformed into a step-by-step regression process. This regression task is defined as follows: (Formula 7) (Formula 8) in, and These represent the start and end indices of the motion sequence, Indicates the length of the historical human movement sequence. This represents the output motion sequence of the previous regression step. This represents the progressive decoding function at step k. Represents global translation and This indicates the global orientation. Through this progressive regression approach, the model can continuously utilize coarse-grained global trajectory and local scene information at each prediction stage to gradually correct future motion, thereby improving the stability and physical plausibility of long-term motion prediction.

[0020] 3-3. Motion Encoding and Interactive Scene Feature Fusion: In the k-th regression step, the MLP-based motion encoder processes the motion sequence. Extracting motion features This characteristic is consistent with the characteristics of self-centered interaction scenarios. Fusion, through a Transformer-based pose decoder Predict the next motion sequence: (Formula 9) 3-4. Traverse the complete motion sequence using a sliding window. Next, the motion sequences obtained from each regression step are input into a GRU-based motion decoder for refinement, smoothing the transition between adjacent regression steps and reducing accumulated errors in long sequence predictions. Subsequently, a pre-trained VPoser model is used to process the motion sequences. Decode and constrain to obtain posture parameters that conform to the prior human posture; input the posture parameters, global translation and global orientation into the SMPL-X parameterized human model to obtain the corresponding three-dimensional human motion.

[0021] Furthermore, the present invention also includes a three-dimensional human motion prediction system with global navigation and local interaction, comprising: Sparse global trajectory planning module: operates on sparse scene representation, constructs a global occupancy graph as a complementary geometric prior, combines a gaze-guided multimodal fusion strategy, and uses a Transformer-based trajectory decoder to predict coarse-grained future global trajectories. Adaptive egocentric scene extraction module: Based on the predicted global translation and the instantaneous velocity of the human body, the module dynamically adjusts the human-centered cylindrical perception radius using a velocity-driven scaling strategy, extracts motion-related local sub-regions from the dense global point cloud, and obtains egocentric scene features through a dual-branch coding scheme. Progressive motion decoding module: Introduces variable time step, autoregressively fuses motion features and local scene features through sliding window, uses pose decoder to refine coarse trajectory into high-fidelity motion sequence, and decodes it into 3D human motion by pre-trained model after sequence smoothing.

[0022] Furthermore, the present invention also provides a three-dimensional human motion prediction device with global navigation and local interaction, the device having a built-in three-dimensional human motion prediction system with global navigation and local interaction.

[0023] Beneficial effects of this invention: This invention addresses the challenge of balancing global navigation accuracy and local interaction details in 3D human motion prediction by proposing a hierarchical framework for predicting human motion in 3D scenes. The method employs a sparse global trajectory planner to obtain a coarse-grained global path, an adaptive egocentric scene extractor to dynamically extract motion-related local geometric information from a dense point cloud, and a progressive motion decoder to gradually generate future human motion sequences. Compared to methods relying solely on sparse or complete global scene modeling, this invention reduces computational overhead and suppresses irrelevant spatial noise while preserving the boundaries of key obstacles and interaction geometric details around the human body, thereby improving the accuracy of predicted trajectories, the naturalness of motion postures, and the physical plausibility of human-environment interactions. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall framework of the method of the present invention.

[0025] Figure 2 It is the sparse global trajectory planner module constructed in the method of this invention.

[0026] Figure 3 It is the adaptive egocentric scene extractor module constructed in the method of this invention.

[0027] Figure 4 It is the progressive motion decoding module constructed in the method of this invention. Detailed Implementation

[0028] The detailed parameters of the present invention will be further described below.

[0029] like Figure 1 As shown, this invention provides a three-dimensional human motion prediction method based on global navigation and local interaction, with the following specific steps: Step (1) Construct a Sparse Global Trajectory Planner (SGTP), such as Figure 2 As shown.

[0030] 1-1. To provide efficient global trajectory prediction guidance, SGTP is used in sparse scene representation. The above operation. Since sparse scene representations inevitably discard fine geometric and structural details, a global occupancy graph is explicitly constructed. As a complementary geometric prior, the global scene is discretized into a two-dimensional grid of resolution H×W. In this embodiment, H=W=128. A representative height is estimated for each grid block. Calculated according to Formula 1. In this formula, when a point exists within the block ( ), take the maximum z-coordinate of these points as the height value; when the block is empty ( When filling a grid cell, the average height of surrounding blocks is used for smooth spatial filling. For each grid cell, a representative height is estimated. It is used to characterize the geometric properties in the vertical direction. The formula for calculating a height map is: When the block (i,j) contains the point (i.e.) ): (Formula 1) When block (i,j) is empty (i.e. ), using neighborhood smoothing: (Formula 2) in, Let (i,j) represent the set of points that fall within block (i,j). This represents the set of blocks surrounding it in a two-dimensional grid. The formula applies to empty blocks ( Apply local spatial smoothing to fill missing values.

[0031] 1-2. Occupation Graph Construction: Deriving Binary Occupation States Based on the Heightmap Occupation map is processed by MLP to extract occupancy features. .

[0032] (Formula 3) in, For indicator functions, when Exceeding the threshold At that time, it will occupy the state. Set to 1. Construct the occupancy graph. Subsequently, MLP processing was used to extract occupancy features. .

[0033] 1-3. Motion Feature Encoding and Fusion: To fully utilize the correlation between historical human motion states and prior global scene occupancy, and to enhance the model's joint perception of human motion trends, passable areas, and potential obstacle areas, the input motion sequence is encoded and fused. Encoded as raw motion embedding, and The final motion characteristics are obtained by fusion. At the same time, sparse scenes Encoding as global sparse scene features The number of points in the sparse scene is set to 4096.

[0034] 1-4. Gaze-Guided Interaction Perception: To transform human gaze information into spatial guidance signals for potential interactive targets and enhance the model's ability to recognize target objects and high-interaction-probability areas, gaze-guided interaction is achieved through scene features guided by gaze point G. Calculate attention to obtain the target region representation. The standard multimodal fusion strategy is used for aggregation. , and To obtain interactive scene features Highlight areas with a high probability of interaction.

[0035] 1-5. Trajectory Prediction: Using a Transformer-based trajectory decoder Predict the complete global trajectory. The decoder uses... and Given the input, predict the future global orientation Õ and translation P̃. In this embodiment, the trajectory decoder is implemented as a single-layer Transformer with 4 attention heads and a dimension of 256, predicting the future global orientation Õ and translation P̃: (Formula 4) Step (2) Construct an adaptive egocentric scene extractor, such as Figure 3 As shown, its core purpose is to filter out irrelevant scene noise far from the human movement path, reduce the computational overhead caused by dense point cloud processing, and at the same time preserve key local geometric details such as obstacle boundaries, contact surfaces and passable areas, so as to provide high-fidelity, human-centered local scene constraints for subsequent progressive motion decoding, thereby improving the physical rationality and scene consistency of predicted motion.

[0036] The scene extraction module dynamically crops high-density human-centered sub-scenes and adaptively adjusts the perception radius based on instantaneous velocity to suppress global spatial noise and preserve key local geometric priors, as detailed below: 2-1. First, given the estimated global translation of step (1). Define a basic cylinder with radius... The geometry naturally aligns with the spatial dimensions of the human body. At time step k, the cylinder undergoes a current global translation. Centered on the target. Calculate the instantaneous velocity. And according to speed The cylinder size is dynamically adjusted. The adaptive scaling strategy is defined as follows: (Formula 5) (Formula 6) Where λ is a hyperparameter controlling the radius expansion. Extracted subset Points satisfying the following conditions Composition: The horizontal distance from the center of the cylinder does not exceed And the vertical coordinate is subject to a height threshold. limit, . , These represent the x and y coordinates at the current k-th time step, respectively. This represents a dense global point cloud.

[0037] 2-2. From the extracted subset Uniform sampling Standardized local scenarios are constructed from individual points. In this embodiment, Dense scenes A dual-branch coding scheme is adopted: (1) geometric prior coding, including two complementary priors: height constraint coding and human proximity coding; (2) scene encoder extracts local geometric features. After aggregation, an egocentric scene representation is formed. .

[0038] Step (3) Construct a progressive motion decoder. For example... Figure 4 As shown, this module takes historical human motion sequences, global translation and orientation trajectories output by SGTP, and local scene features output by AESE as input. Through last-frame replication, trajectory alignment, sliding windowing, and autoregressive prediction, it gradually generates future motion. Its core objective is to reduce the error accumulation and attitude drift caused by one-time prediction of long sequences, ensuring that the prediction results simultaneously satisfy global navigation orientation and local geometric constraints. The specific steps are as follows: 3-1. The progressive motion decoder copies the last frame of the historical human motion sequence, thus converting the historical human motion sequence into a single sequence. Time fill to To obtain the initial motion sequence Then fill the sequence. By aligning with the coarse-grained global trajectory and updating the global translation and global orientation, a motion sequence aligned with the coarse-grained global trajectory is obtained. .

[0039] 3-2. To avoid error accumulation and motion drift when predicting long motion sequences in a single step, and to enable the model to progressively combine the motion results and scene constraint information from the previous stage during the prediction process, a variable time step *u* is introduced to control the length of each new prediction sequence, transforming the future motion prediction process into a step-by-step regression process. This regression task is defined as follows: (Formula 7) (Formula 8) in, and These represent the start and end indices of the motion sequence, Indicates the length of the historical human movement sequence. This represents the output motion sequence of the previous regression step. This represents the progressive decoding function at step k. Represents global translation and This indicates the global orientation. Through this progressive regression approach, the model can continuously utilize coarse-grained global trajectory and local scene information at each prediction stage to gradually correct future motion, thereby improving the stability and physical plausibility of long-term motion prediction.

[0040] 3-3. Motion Encoding and Interactive Scene Feature Fusion: In the k-th regression step, the MLP-based motion encoder processes the motion sequence. Extracting motion features This characteristic is consistent with the characteristics of self-centered interaction scenarios. Fusion, through a Transformer-based pose decoder Predict the next motion sequence: (Formula 9) 3-4. Traverse the complete motion sequence using a sliding window. Next, the motion sequences obtained from each regression step are input into a GRU-based motion decoder for refinement, smoothing the transition between adjacent regression steps and reducing accumulated errors in long sequence predictions. Subsequently, a pre-trained VPoser model is used to process the motion sequences. Decode and constrain to obtain posture parameters that conform to the prior human posture; input the posture parameters, global translation and global orientation into the SMPL-X parameterized human model to obtain the corresponding three-dimensional human motion.

[0041] All experiments were performed on a single NVIDIA GeForce RTX 3090 GPU. SGTP was pre-trained to mitigate initial bias. The framework was jointly optimized using AdamW with an initial learning rate of [missing information]. Weight decay Exponential decay scheduling (γ=0.98) was used. Both the motion encoder and gaze encoder were instantiated as 3-layer Transformers with 8 attention heads and 256-dimensional hidden states, mapping features to a 256-dimensional latent space. A Dropout rate of 0.3 was applied to the network.

[0042] Experimental results on the GIMO dataset show that Aegis achieves the best results in terms of trajectory error (average 472mm / final 532mm) and a success rate of 56.7%, both significantly outperforming all baseline methods. Specific comparisons are as follows: Table 1: Comparison Table of Multiple Model Experiments

[0043] As shown in Table 1, the method of this invention achieves the best trajectory prediction performance and the highest navigation success rate on the GIMO dataset. Compared with existing methods such as AuxFormer, ContAware, BiFU, HUMOF, SIF3D, and GAP3DS, the average trajectory error and final trajectory error of the method of this invention are reduced to 472mm and 532mm, respectively, while the success rate is improved to 56.7%. These results demonstrate that this invention obtains stable global navigation priors through a sparse global trajectory planner, preserves motion-related local geometric details through an adaptive egocentric scene extractor, and progressively refines future human motion sequences through a progressive motion decoder. This effectively reduces trajectory drift during long-term predictions and improves the physical consistency between human motion and the 3D scene. Furthermore, integrating the AESE module into existing methods such as HUMOF, SIF3D, and GAP3DS reduces the trajectory error and improves the success rate of these methods, indicating that the module has good versatility and scalability.

[0044] This application also provides an electronic device, including a processor and a memory.

[0045] The memory is used to store computer programs.

[0046] When the processor executes a program stored in the memory, it implements any of the methods described in this application.

[0047] In one possible implementation, the electronic device of this application embodiment further includes a communication interface and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0048] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0049] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0050] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0051] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0052] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements any of the methods described in this application.

[0053] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the methods described in this application.

[0054] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0055] 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0056] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A three-dimensional human motion prediction method based on global navigation and local interaction, characterized in that, include: Step (1) Construct a sparse global trajectory planner: Construct a global occupancy map based on sparse scene point cloud, use the global occupancy map as a global geometric prior, and combine historical human motion sequences, sparse scene features and gaze point information to predict coarse-grained global trajectories in future time steps. Step (2): Construct an adaptive egocentric scene extractor: Determine the estimated position of the human body in the future time step based on the coarse-grained global trajectory, and dynamically adjust the size of the local perception area centered on the human body based on the instantaneous velocity of the human body. Extract local sub-scenes related to the current motion from the dense global scene point cloud to obtain the egocentric local scene features corresponding to each local sub-scene. Step (3) Construct a progressive motion decoder: Align the filling sequence with the coarse-grained global trajectory, and fuse motion features and local scene features within a sliding window. Use an autoregressive approach to gradually predict future human motion sequences, and stitch the future human motion sequences with historical human motion sequences to finally generate a 3D human motion that conforms to global path navigation constraints and local scene geometric constraints.

2. The three-dimensional human motion prediction method based on global navigation and local interaction according to claim 1, characterized in that, Step (1) Discretize the sparse scene point cloud into a two-dimensional grid, estimate the representative height of each grid block to characterize the geometric structure of the grid block in the vertical direction; when there are scene points in the grid block, take the maximum height of the scene point in the grid block as the representative height of the grid block; when the grid block is empty, perform spatial smooth filling based on the representative heights of its four neighboring grid blocks; determine the occupancy status of each grid block based on the comparison result of the representative height and the preset height threshold, and obtain the global occupancy map; When there is a scene point in the grid block (i,j), that is : (Official 1) When the grid block (i,j) is empty, i.e. Utilizing neighborhood smoothing: (Official 2) in, This represents the set of points that fall within grid block (i,j); Represents the set of surrounding grid blocks of grid block (i,j); Represents the neighborhood grid block The representative height; Global Occupancy Map Derived Based on Representative Height : (Official 3) in, As an indicator function, when the representative height Exceeding the threshold At that time, the global occupation map will be... The occupied state is set to 1.

3. The three-dimensional human motion prediction method based on global navigation and local interaction according to claim 2, characterized in that, Step (1) also includes: The constructed global occupancy graph The input is then fed into a multilayer perceptron for feature extraction, extracting the dominant features. ; The input historical human motion sequence is encoded into a raw motion embedding and then combined with the occupation feature. Fusion is performed to obtain motion characteristics; The sparse scene point cloud is encoded into global sparse scene features, and attention computation is performed on the global sparse scene features based on the gaze point information to obtain the target region representation; Multimodal fusion of motion features, global sparse scene features, and target region representation is performed to obtain interactive scene features; By inputting motion features and interaction scene features into a Transformer-based trajectory decoder, global translation and global orientation can be predicted within a future time period.

4. The three-dimensional human motion prediction method based on global navigation and local interaction according to claim 1, characterized in that, The adaptive egocentric scene extractor described in step (2) employs a velocity-driven local perception region scaling strategy. The local perception region is a cylindrical region centered on the estimated position of the human body, and its radius is dynamically determined based on the instantaneous velocity of the human body. ； ； in, This represents the local sensing radius at the k-th time step. Indicates the basic sensing radius. Indicates the radius expansion coefficient. This represents the instantaneous velocity of the human body at the k-th time step; Hyperparameters for controlling radius expansion; extracted subsets Points satisfying the following conditions Composition: The horizontal distance from the center of the cylindrical region does not exceed And the vertical coordinate is subject to a height threshold. limit, ; , These represent the x and y coordinates at the current k-th time step, respectively. This represents a dense global point cloud.

5. The three-dimensional human motion prediction method based on global navigation and local interaction according to claim 4, characterized in that, Step (2) adopts a dual-branch coding scheme, specifically as follows: From the extracted subset Uniform sampling Each point is used to construct a standardized local scene. The standardized local scene is feature-encoded using a dual-branch coding scheme. The geometric prior coding branch enhances point features using height-constrained coding and human proximity coding. The feature extraction branch extracts local geometric features using a scene encoder. The features output from both branches are then aggregated to obtain the time step. Corresponding self-centered scenario representation .

6. The three-dimensional human motion prediction method based on global navigation and local interaction according to claim 5, characterized in that, The geometric prior coding branch constructs geometric prior features based on the spatial geometric relationships of local scene points. Among them, the height constraint coding is used to describe the vertical spatial relationship of scene points relative to the ground, and the human proximity coding is used to characterize the distance relationship between scene points and the center of the human body, thereby enhancing the ability of local scene features to express information related to human motion interaction. The feature extraction branch extracts point-level geometric features of the normalized local scene through the scene encoder.

7. The three-dimensional human motion prediction method based on global navigation and local interaction according to claim 1, characterized in that, The progressive regression described in step (3) specifically refers to: 3-1. The progressive motion decoder copies the last frame of the historical human motion sequence, thus transforming the historical human motion sequence... Time fill to Then fill the sequence. By aligning with the coarse-grained global trajectory and updating the global translation and global orientation, a motion sequence aligned with the coarse-grained global trajectory is obtained. ; 3-2. Introducing a variable time step By controlling the length of each new prediction sequence, the future motion prediction process is transformed into a step-by-step regression process; 3-3. Motion Encoding and Interactive Scene Feature Fusion: In the k-th regression step, the MLP-based motion encoder processes the motion sequence. Extracting motion features This feature is consistent with the characteristics of egocentric interaction scenarios. Fusion, using an attitude decoder to predict the next motion sequence; 3-4. Traverse the complete motion sequence using a sliding window. Then, the motion sequences obtained from each regression step are input into a GRU-based motion decoder for refinement; a pre-trained VPoser model is then used to process the motion sequences. Decode and constrain to obtain posture parameters that conform to the prior human posture; input the posture parameters, global translation and global orientation into the SMPL-X parameterized human model to obtain the corresponding three-dimensional human motion.

8. A three-dimensional human motion prediction system with global navigation and local interaction, characterized in that, The system is used to implement the method as described in claim 1, comprising: Sparse global trajectory planning module: operates on sparse scene representation, constructs a global occupancy graph as a complementary geometric prior, combines a gaze-guided multimodal fusion strategy, and uses a Transformer-based trajectory decoder to predict coarse-grained future global trajectories. Adaptive egocentric scene extraction module: Based on the predicted global translation and the instantaneous velocity of the human body, the module dynamically adjusts the human-centered cylindrical perception radius using a velocity-driven scaling strategy, extracts motion-related local sub-regions from the dense global point cloud, and obtains egocentric scene features through a dual-branch coding scheme. Progressive motion decoding module: Introduces variable time step, autoregressively fuses motion features and local scene features through sliding window, uses pose decoder to refine coarse trajectory into high-fidelity motion sequence, and decodes it into 3D human motion by pre-trained model after sequence smoothing.

9. A three-dimensional human motion prediction device with global navigation and local interaction, characterized in that, The device includes the system as described in claim 9.

Images