A method for reconstructing four-dimensional Gaussian dynamic scenes based on single-exposure compressed imaging
By employing a four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging, and utilizing deep learning and optimization iterative algorithms, dynamic scenes can be reconstructed under single-exposure conditions. This solves the dynamic modeling problem under single-exposure compressed observation and achieves efficient and accurate dynamic scene reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WESTLAKE UNIV
- Filing Date
- 2026-04-20
- Publication Date
- 2026-05-29
AI Technical Summary
Existing dynamic scene reconstruction methods based on 4D Gaussian representation struggle to achieve high-quality reconstruction under single-exposure compressed observation conditions. This is mainly because multi-frame information is encoded and superimposed into single-frame measurement data, making the time dimension no longer explicitly separable. As a result, dynamic information becomes a latent variable, and existing methods have failed to effectively solve the dynamic modeling problem under single-exposure compressed observation conditions.
A four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging is adopted. By acquiring single-exposure compressed images and decoding them using an optimized iterative algorithm, combined with motion recovery and skeleton generation algorithms from deep learning, three-dimensional Gaussian parameters are constructed and rendered. The reconstruction results are optimized using a photometric loss function to achieve high-quality reconstruction of dynamic scenes.
While reducing data acquisition and transmission costs, it achieves high-quality dynamic scene reconstruction under single-exposure compression conditions, improves reconstruction accuracy and efficiency in high-speed motion and bandwidth-limited environments, conforms to the real imaging physical process, and improves the reliability of reconstruction results.
Smart Images
Figure CN122115741A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computational imaging and 3D vision reconstruction technology, and in particular to a method for reconstructing a four-dimensional Gaussian dynamic scene based on single-exposure compressed imaging. Background Technology
[0002] Reconstructing dynamic scenes from images is a crucial research problem in computer vision, with wide applications in virtual reality, augmented reality, intelligent manufacturing, autonomous driving, and robot perception, providing key technological support for motion analysis and dynamic scene understanding. Existing dynamic scene reconstruction methods based on 4D Gaussian representation typically rely on complete temporal observation data, i.e., modeling dynamic scenes using continuous multi-frame images or multi-view time-series data. However, in high-speed motion scenes, low-light environments, and bandwidth-constrained transmission systems or embedded terminal devices, limitations such as sensor frame rate, storage capacity, and data transmission rate often make it difficult to obtain complete multi-frame temporal data, thus restricting the application of these methods in real-world scenarios.
[0003] To improve the efficiency of spatiotemporal information acquisition, Snapshot Compressive Imaging (SCI) uses a time modulation coding mechanism to compress and superimpose information from multiple time frames into the measurement results of a single exposure, achieving compressed acquisition of high-dimensional spatiotemporal information. This has significant application value in high-speed imaging and low-bandwidth data acquisition scenarios. For example, Chinese patent document CN117761896A discloses a single-exposure compressed ultrafast coherent modulation imaging device; Chinese patent document CN116366993A discloses a method and imaging system for acquiring four-dimensional compressed images from a single exposure.
[0004] However, under single-exposure compressed observation conditions, information from multiple time frames is encoded and superimposed into a single frame of measurement data. The time dimension is no longer explicitly separable, and the original dynamic information becomes a latent variable. This transforms the dynamic scene reconstruction problem from a representation optimization problem under complete observation conditions into an underdetermined compression inverse problem. Existing dynamic scene reconstruction methods based on 4D Gaussian representations are all based on the assumption of complete temporal observation and have not yet solved the dynamic modeling problem under single-exposure compressed observation conditions.
[0005] Therefore, there is an urgent need for a technical solution that can realize 4D Gaussian modeling and reconstruction of dynamic scenes under single-exposure compressed observation conditions, so as to achieve high-quality dynamic scene reconstruction while reducing data acquisition and transmission costs. Summary of the Invention
[0006] This invention provides a four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging, which can achieve high-quality dynamic scene reconstruction while reducing data acquisition and transmission costs.
[0007] A method for four-dimensional Gaussian dynamic scene reconstruction based on single-exposure compressed imaging includes: Step 1: Obtain a single single-exposure compressed image and its corresponding coded mask, and decode the single-exposure compressed image based on an optimized iterative algorithm to reconstruct a multi-view image sequence; Step 2: Use a deep learning-based structure-of-motion (SOG) algorithm to process the multi-view image sequence and estimate the initial scene point cloud and camera pose parameters. Step 3: Process the multi-view image sequence using a deep learning-based skeleton generation algorithm to obtain the initial scene skeleton structure; Step 4: Initialize the 3D Gaussian parameters based on the initial scene point cloud and camera pose parameters to construct a normalized spatial 3D Gaussian representation; Step 5: Construct a skeleton deformation field based on the initial scene skeleton structure, and perform skeleton-driven deformation on the three-dimensional Gaussian in the normal space using the linear hybrid skinning method to obtain the deformed Gaussian representation; Step 6: Render the deformed Gaussian representation based on the differentiable rendering method to obtain rendered images corresponding to different times and viewpoints; Step 7: Input the rendered image into the single-exposure compression imaging model to generate the synthesized single-exposure measurement value, and perform multiple rounds of iterative optimization and update of the network parameters by constructing a photometric loss function to achieve dynamic scene reconstruction.
[0008] In step 1, the optimization iterative algorithm is a total variational algorithm based on generalized alternating projection, and it is solved iteratively; in the... In each iteration, the current image estimate is updated based on the auxiliary variables obtained from the previous iteration, and the updated image estimate is used to update the auxiliary variables in order to gradually approach the optimal solution.
[0009] In step 2, the deep learning-based structure-of-motion (SOP) algorithm is the VGGSfM algorithm, and the formula is as follows: ; in, Represents a sequence of multi-view images. Indicates camera pose parameters. Representing scene point clouds, This represents the VGGSfM algorithm.
[0010] In step 3, the deep learning-based skeleton generation algorithm is the RigAnything algorithm, and the formula is as follows: ; in, Represents a sequence of multi-view images. This represents the initial skeleton structure diagram. This represents the RigAnything algorithm.
[0011] In step 4, the point cloud data obtained in step 2 is initialized with 3D Gaussian parameters to construct a normalized spatial 3D Gaussian representation. The scene is modeled using 3D Gaussian Splatting (3DGS), which uses a set of Gaussian primitives. To represent a scene, where each Gaussian element... By Gauss Center , by quaternions Representing rotation parameters and scaling vectors Opacity parameter and spherical harmonic coefficients used to describe view-dependent color information To be determined jointly.
[0012] In step 5, a skeleton deformation field is constructed based on the initial scene skeleton structure. The specific process is as follows: The skeleton deformation field is modeled using a multilayer perceptron, and its expression is as follows: ; in, Indicates the first Each joint in time Local spinor, Indicates the first Each joint in time The local translation amount, Indicates position code, Represents the deformation field of the skeleton; Furthermore, the global transformations corresponding to each joint are obtained through forward kinematics calculations, and their expressions are as follows: ; in, Indicates the first Each joint in time Global rotation amount, Indicates the first Each joint in time Global translation amount This represents the initial skeleton structure diagram. This represents a forward kinematics operation, used to propagate the local transformations of each joint to its sub-joints according to the skeleton topology, thereby obtaining the global transformations of each joint.
[0013] In step 5, the skeleton-driven deformation of the three-dimensional Gaussian in the normal space is performed using the linear mixture skinning method to obtain the deformed Gaussian representation. The specific process is as follows: First, build Each bone has 10 bones. Corresponding to the connecting joint The bone border between its parent joint and the joint; Subsequently, each Gaussian center in the static state of the gauge space is... The skeleton-driven deformation is mapped to time. The position is used to obtain the corresponding dynamic Gaussian representation: ; in, Represents learnable skin weights, satisfying , express The center of Gauss at time; Learnable skin weights are generated using a skin correction field; wherein, the skin correction field is modeled using a multilayer perceptron, and its expression is as follows: ; ; ; in, Describing the center of regular Gaussian and bones The distance between them It is every bone The learnable radius of influence Indicates the skin correction field. This indicates the position code.
[0014] The specific process of step 6 is as follows: Perform differential Gaussian rasterization on the deformed Gaussian representations corresponding to different timestamps obtained in step 5 to generate rendered images corresponding to different times and viewpoints, as shown in the following expression: ; in, This represents the generated rendered image. This represents the Gaussian index sorted by depth along the line of sight. This represents the view-dependent color value calculated from the spherical harmonic coefficients. Indicates opacity The weighted result of the two-dimensional Gaussian distribution obtained by projecting the three-dimensional Gaussian distribution onto the two-dimensional imaging plane.
[0015] In step 7, the rendered image is input into the single-exposure compression imaging model to generate a synthesized single-exposure measurement value, expressed as follows: ; in, This represents the composite single-exposure compression measurement. It is the first Rendered image, This represents the corresponding modulation mask.
[0016] In step 7, the photometric loss function is constructed, which includes an absolute value error loss term and a structural similarity loss term, as shown in the following expression: ; in, Synthesized exposure compression measurements Compared with the actual single-exposure compression measurement value The photometric loss function between them For the absolute value error loss term, For structural similarity loss term, This is a coefficient used to adjust the weight ratio of the two loss terms.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces a single-exposure compressed imaging model to encode information from multiple time frames into single-exposure measurement data, enabling the acquisition of dynamic scene information without the need for continuous multi-frame acquisition. This effectively reduces data acquisition costs and storage overhead, and improves the applicability of the system in high-speed motion scenes and bandwidth-constrained environments.
[0018] 2. This invention utilizes the total variational algorithm based on generalized alternating projection (GAP-TV) to reconstruct multi-view image sequences from single-exposure compressed observation data, and combines it with deep learning motion recovery structure method to estimate scene point clouds and camera pose, thereby achieving three-dimensional reconstruction of dynamic scenes under single-exposure observation conditions.
[0019] 3. This invention obtains an initial skeleton structure by constructing a skeleton generation algorithm, and introduces a skeleton deformation field and a linear hybrid skin model to drive the skeleton deformation of three-dimensional Gaussian elements, so that the structural motion of dynamic targets can be modeled through skeleton constraints, thereby improving the accuracy and stability of complex dynamic scene reconstruction.
[0020] 4. This invention uses three-dimensional Gaussian scattering as a scene representation method and optimizes it by combining a differentiable rendering mechanism. Compared with dynamic reconstruction methods based on implicit neural representation, it can significantly improve rendering efficiency and training speed while ensuring reconstruction quality.
[0021] 5. This invention incorporates a single-exposure compressed imaging physical model into the optimization process, aligning the rendered results with the actual compressed measurement data through photometric consistency constraints. This makes the dynamic scene reconstruction results more consistent with the real imaging physical process, improving the reliability of the reconstruction results. Therefore, this invention achieves effective modeling and reconstruction of dynamic scenes under compressed observation conditions, providing a new technical approach for low-bandwidth, high-efficiency dynamic scene reconstruction. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart of a four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging, according to an embodiment of the present invention.
[0024] Figure 2 This is a schematic diagram of the single exposure compression imaging (SCI) model in this invention.
[0025] Figure 3 This is a schematic diagram of the principle of four-dimensional Gaussian dynamic scene reconstruction based on single-exposure compressed imaging according to the present invention.
[0026] Figure 4 This is a comparison of the dynamic scene image reconstruction effects of single-exposure compression imaging according to an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.
[0028] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.
[0029] like Figure 1 As shown, a four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging includes the following steps: Step 1: Obtain a single-exposure compressed measurement image.
[0030] In one exemplary embodiment, taking a human posture motion scene as an example, compressed measurement data of the dynamic scene is acquired using a single-exposure compressed imaging (SCI) system. Figure 3 As shown, dynamic scenes form continuous time-series images during the acquisition process. Each time frame image is associated with its corresponding modulation mask. Pixel-by-pixel modulation calculations are performed and accumulated over the exposure time to obtain a single-exposure compressed measurement image. As shown in the following formula: ; in, This represents the compression measurement value for a single exposure. It is the first Original image frame Indicates the first Frame modulation mask, This indicates a pixel-by-pixel multiplication operation.
[0031] Step 2: Using the single-exposure compressed image and its corresponding coded mask obtained in Step 1 as input, the SCI image is decoded using a physics-based model-driven decoding algorithm to reconstruct the corresponding multi-view image sequence.
[0032] In one exemplary embodiment, such as Figure 2 As shown, a physically-based model-driven decoding algorithm is used to process the acquired single-exposure compressed image. In this embodiment, the decoding algorithm can be a physically-based model-driven algorithm such as GAP-TV, and its mathematical model is as follows: ; in, This represents a sequence of multi-view images to be reconstructed. This represents the input compression measurement value. Represents the sensing matrix, For the introduced auxiliary variables, Here is the regularization parameter; the GAP-TV algorithm solves iteratively, at the th In each iteration, the current image estimate is updated based on the auxiliary variables obtained from the previous iteration, and the updated image estimate is used to update the auxiliary variables in order to gradually approach the optimal solution.
[0033] Step 3: Input the multi-view image sequence obtained in Step 2 into the learning-based structure recovery module to perform initial 3D structure estimation of the scene, thereby obtaining the initial point cloud and camera pose parameters of the scene, which are then used for the initialization of 3D Gaussian parameters.
[0034] In one exemplary embodiment, such as Figure 2 As shown, a deep learning-based structure reconstruction algorithm is used to process the multi-view image sequence obtained in step 1. In this embodiment, the structure reconstruction algorithm can be a learning-based structure reconstruction algorithm such as VGGSfM. By extracting and matching features from the multi-view image sequence, camera pose parameters and scene point clouds are obtained, and its representation is shown in the following formula: ; in, Represents a sequence of multi-view images. Indicates camera pose parameters. Representing scene point clouds, This represents the VGGSfM algorithm.
[0035] Then, the scene point cloud data is initialized with 3D Gaussian parameters to construct a normalized spatial 3D Gaussian representation; the scene is modeled using 3D Gaussian Splatting (3DGS), which uses a set of Gaussian primitives. To represent a scene, where each Gaussian element... By Gauss Center , by quaternions Representing rotation parameters and scaling vectors Opacity parameter and spherical harmonic coefficients used to describe view-dependent color information To be determined jointly.
[0036] Step 4: Input the multi-view image sequence obtained in Step 2 into the learning-based skeleton generation module to perform initial skeleton structure estimation for the scene.
[0037] In one exemplary embodiment, such as Figure 2 As shown, a deep learning-based skeleton generation algorithm is used to process the multi-view image sequence obtained in step 2. In this embodiment, the skeleton generation algorithm can employ a learning-based skeleton generation model such as RigAnything. By extracting features and inferring structure from the multi-view image sequence, it predicts the skeleton structure parameters of the scene. Its mathematical expression is as follows: ; in, Represents a sequence of multi-view images. This represents the initial skeleton structure diagram. This represents the RigAnything algorithm.
[0038] Step 5: Input the initial scene skeleton structure obtained in Step 4 into the skeleton deformation field, and perform skeleton-driven deformation on the normal space 3D Gaussian using the linear blending skin method to obtain the deformed Gaussian representation.
[0039] In one exemplary embodiment, such as Figure 2 As shown, the initial scene skeleton structure obtained in step 3 is subjected to a skeleton deformation field. Attitude prediction at any given time. In this embodiment, the skeleton deformation field is modeled using a multilayer perceptron, and its expression is as follows: ; in, Indicates the first Each joint in time Local spinor, Indicates the first Each joint in time The local translation amount, Indicates position code, This represents the deformation field of the skeleton.
[0040] Then, the global transformations corresponding to each joint are obtained through forward kinematics calculations, and their expressions are as follows: ; in, Indicates the first Each joint in time Global rotation amount, Indicates the first Each joint in time Global translation amount This represents the initial skeleton structure diagram. This represents a forward kinematics operation, used to propagate the local transformations of each joint to its sub-joints according to the skeleton topology, thereby obtaining the global transformations of each joint.
[0041] After obtaining the joint pose, a Gaussian deformation field is constructed using a learnable linear blend skinning (LBS) deformation model. First, a Gaussian deformation field is constructed. Each bone has 10 bones. Corresponding to the connecting joint The bone edges between its parent joints; subsequently, each Gaussian center in the static state of the normal space. The skeleton-driven deformation is mapped to time. The position is used to obtain the corresponding dynamic Gaussian representation.
[0042] ; in, Represents learnable skin weights, satisfying , express The Gaussian center at a given moment.
[0043] Learnable skin weights are generated using a skin correction field; the skin correction field is modeled using a multilayer perceptron, and its expression is as follows: ; ; ; in, Describing the center of regular Gaussian and bones The distance between them It is every bone The learnable radius of influence Indicates the skin correction field. This indicates the position code.
[0044] Step 6: Input the deformed Gaussian obtained in Step 5 into the differentiable rendering module, and generate rendering images corresponding to different times and viewpoints through rendering calculations.
[0045] In one exemplary instance, such as Figure 2 As shown, differential Gaussian rasterization is performed on the deformed Gaussian representations corresponding to different timestamps obtained in step 5 to generate a multi-view image sequence under the corresponding timestamps, the expression of which is as follows: ; in, Represents an image. This represents the Gaussian index sorted by depth along the line of sight. This represents the view-dependent color value calculated from the spherical harmonic coefficients. Indicates opacity The weighted result of the two-dimensional Gaussian distribution obtained by projecting the three-dimensional Gaussian distribution onto the two-dimensional imaging plane.
[0046] Step 7: Input the rendered image obtained in Step 6 into the single-exposure compression imaging model to generate the synthesized single-exposure measurement value, and perform multiple rounds of iterative optimization and update of the network parameters through photometric consistency constraints to achieve dynamic scene reconstruction.
[0047] In one exemplary instance, such as Figure 2 As shown, the multi-view image sequence obtained in step 6 is input into the single-exposure compression imaging model to generate synthesized single-exposure measurements, the expression of which is as follows: ; in, This represents the composite single-exposure compression measurement. It is the first Rendered image, This represents the corresponding modulation mask.
[0048] Constructing synthetic measurements Compared with the actual single-exposure compression measurement value Photometric loss function between The Gaussian deformation field parameters are iteratively updated and optimized through backpropagation, as shown in the following equation: ; in, For the absolute value error loss term, For structural similarity loss term, In this embodiment, the coefficient used to adjust the weight ratio of the two loss terms is... The value is 0.5.
[0049] During model training, the AdamW optimizer was used to update the network parameters, and its default parameter settings were employed. The total number of training iterations was approximately 20,000. Regarding the learning rate setting, the learning rate for the camera pose parameters was set to 1×10⁻⁶. -3 Gradually decays to 1×10 -5 The learning rate for the skeleton deformation field and the skin correction field is 1×10 -3 It gradually decreased to 1.6 × 10 -5 In the 3D Gaussian initialization phase, the initial point cloud is downsampled to approximately 10,000 Gaussian points to reduce computational complexity and improve training efficiency. After training, the network weight parameters obtained during training are saved for subsequent dynamic scene reconstruction and inference.
[0050] like Figure 4 The image shown is a comparison of the dynamic scene image reconstruction results using single-exposure compressed imaging in this embodiment. The first row shows the realistic dynamic RGB image of the robot toy; the second row shows the realistic dynamic RGB image of the robot toy reconstructed using the method of this invention; the third row shows the realistic dynamic RGB image of the Lego toy; and the fourth row shows the realistic dynamic RGB image of the Lego toy reconstructed using the method of this invention. Figure 4 The results show that the four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging proposed in this invention can achieve high image reconstruction quality under different dynamic scenes, verifying the effectiveness and superiority of this method.
[0051] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for reconstructing a four-dimensional Gaussian dynamic scene based on single-exposure compressed imaging, characterized in that, include: Step 1: Obtain a single single-exposure compressed image and its corresponding coded mask, and decode the single-exposure compressed image based on an optimized iterative algorithm to reconstruct a multi-view image sequence; Step 2: Use a deep learning-based structure-of-motion (SOG) algorithm to process the multi-view image sequence and estimate the initial scene point cloud and camera pose parameters. Step 3: Process the multi-view image sequence using a deep learning-based skeleton generation algorithm to obtain the initial scene skeleton structure; Step 4: Initialize the 3D Gaussian parameters based on the initial scene point cloud and camera pose parameters to construct a normalized spatial 3D Gaussian representation; Step 5: Construct a skeleton deformation field based on the initial scene skeleton structure, and perform skeleton-driven deformation on the three-dimensional Gaussian in the normal space using the linear hybrid skinning method to obtain the deformed Gaussian representation; Step 6: Render the deformed Gaussian representation based on the differentiable rendering method to obtain rendered images corresponding to different times and viewpoints; Step 7: Input the rendered image into the single-exposure compression imaging model to generate the synthesized single-exposure measurement value, and perform multiple rounds of iterative optimization and update of the network parameters by constructing a photometric loss function to achieve dynamic scene reconstruction.
2. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, In step 1, the optimization iterative algorithm is a total variational algorithm based on generalized alternating projection, and it is solved iteratively; in the... In each iteration, the current image estimate is updated based on the auxiliary variables obtained from the previous iteration, and the updated image estimate is used to update the auxiliary variables in order to gradually approach the optimal solution.
3. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, In step 2, the motion recovery structure algorithm based on deep learning is the VGGSfM algorithm.
4. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, In step 3, the deep learning-based skeleton generation algorithm is the RigAnything algorithm.
5. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, In step 5, a skeleton deformation field is constructed based on the initial scene skeleton structure. The specific process is as follows: The skeleton deformation field is modeled using a multilayer perceptron, and its expression is as follows: ； in, Indicates the first Each joint in time Local spinor, Indicates the first Each joint in time The local translation amount, Indicates position code, Represents the deformation field of the skeleton; Furthermore, the global transformations corresponding to each joint are obtained through forward kinematics calculations, and their expressions are as follows: ； in, Indicates the first Each joint in time Global rotation amount, Indicates the first Each joint in time Global translation amount This represents the initial skeleton structure diagram. This represents a forward kinematics operation, used to propagate the local transformations of each joint to its sub-joints according to the skeleton topology, thereby obtaining the global transformations of each joint.
6. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, In step 5, the skeleton-driven deformation of the three-dimensional Gaussian in the normal space is performed using the linear mixture skinning method to obtain the deformed Gaussian representation. The specific process is as follows: First, build Each bone has 10 bones. Corresponding to the connecting joint The bone border between its parent joint and the joint; Subsequently, each Gaussian center in the static state of the gauge space is... The skeleton-driven deformation is mapped to time. The position is used to obtain the corresponding dynamic Gaussian representation.
7. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, The specific process of step 6 is as follows: Perform differential Gaussian rasterization on the deformed Gaussian representations corresponding to different timestamps obtained in step 5 to generate rendered images corresponding to different times and viewpoints, as shown in the following expression: ； in, This represents the generated rendered image. This represents the Gaussian index sorted by depth along the line of sight. This represents the view-dependent color value calculated from the spherical harmonic coefficients. Indicates opacity The weighted result of the two-dimensional Gaussian distribution obtained by projecting the three-dimensional Gaussian distribution onto the two-dimensional imaging plane.
8. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging according to claim 1, characterized in that, In step 7, the rendered image is input into the single-exposure compression imaging model to generate a synthesized single-exposure measurement value, expressed as follows: ； in, This represents the composite single-exposure compression measurement. It is the first Rendered image, This represents the corresponding modulation mask.
9. The four-dimensional Gaussian dynamic scene reconstruction method based on single-exposure compressed imaging as described in claim 1, characterized in that, In step 7, the photometric loss function is constructed by including an absolute value error loss term. and structural similarity loss term .