A 3D pose optimization method based on translation and rotation dual-branch temporal filtering

Abstract

The application discloses a 3D pose optimization method based on translation and rotation double-branch time sequence filtering, belongs to the field of three-dimensional target detection and pose estimation, and comprises the following steps: calling a 2D-to-3D promotion module to obtain original 3D pose data of an object, and constructing a parallel processing translation filtering branch and a rotation filtering branch; the translation branch adopts exponential smoothing filtering to perform time sequence smoothing on a translation vector, and the rotation branch performs filtering processing on a rotation matrix by matrix and quaternion conversion and in combination with spherical linear interpolation; a time sequence cache is used to store a previous frame filtering result, a history reference is provided for current frame calculation, the cache is updated for next frame processing, and finally, a filtered translation vector and a rotation matrix are combined to reconstruct and output an optimized 3D bounding box. The application effectively suppresses position jitter and pose jump in time sequence output of a lightweight pose estimation algorithm, improves the stability and accuracy of three-dimensional pose estimation under the premise of maintaining high real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional target detection and attitude estimation, and in particular relates to a 3D attitude optimization method based on translation and rotation dual-branch temporal filtering. Background Technology

[0002] In the field of computer vision, 3D object pose estimation is one of the core technologies for applications such as autonomous driving, augmented reality (AR), and robot navigation. Its purpose is to reconstruct the depth, shape, and pose information of an object from a 2D image. Existing lightweight 3D object detection and pose estimation algorithms typically employ a two-stage processing architecture: the first stage obtains the 2D clipping region of the object through an object detection model, and the second stage uses image clipping to perform 3D bounding box estimation, and calculates the 2D clipping for subsequent frames to reduce the detector's operating frequency. Due to their high inference speed, these algorithms have become the mainstream solution for real-time 3D pose estimation in mobile and embedded devices, capable of outputting raw 3D pose data containing translation vectors and rotation matrices for common rigid bodies such as shoes, chairs, and cups.

[0003] However, the output of the aforementioned lightweight pose estimation algorithms is susceptible to interference from factors such as sensor acquisition errors, changes in ambient lighting, and false detections between algorithm frames. This results in significant temporal jitter and pose jumps in the generated 3D bounding boxes across consecutive video frames, specifically manifested as abrupt changes in translation vector coordinate values ​​and discontinuities in the pose representation of the rotation matrix. Existing optimization methods often employ a single filtering strategy to uniformly process the overall pose data, failing to distinguish between the linear characteristics of the translation vector and the nonlinear rigid body rotation characteristics of the rotation matrix. This makes it difficult to suppress positional jitter while avoiding the destruction of rotational orthogonality, easily leading to pose lag or residual jitter. Furthermore, existing methods lack a temporal state management mechanism designed specifically for the output characteristics of lightweight algorithms, failing to effectively utilize inter-frame pose correlation information and limiting their ability to achieve high stability and high accuracy in 3D pose optimization in real-time applications. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a 3D pose optimization method based on translational and rotational dual-branch temporal filtering, comprising: The 2D to 3D enhancement module is invoked to obtain the original 3D pose data of the object to be optimized in the current frame. The original 3D pose data includes translation vectors and rotation matrices. The translation vector is input into the translation filtering branch, and the filtered translation vector of the current frame is calculated based on the original translation vector of the current frame and the filtered translation vector of the previous frame stored in the translation buffer. The rotation matrix is ​​input into a rotation filtering branch that runs parallel to the translation filtering branch. The rotation matrix is ​​converted into a rotation quaternion for the current frame. Based on the rotation quaternion for the current frame and the previous frame's filtered rotation quaternion stored in the rotation buffer, the current frame's filtered rotation quaternion is calculated, and the current frame's filtered rotation quaternion is converted into a filtered rotation matrix. Update the translation buffer and the rotation buffer with the current frame filtered translation vector and the current frame filtered rotation quaternion, respectively. The filtered translation vector and the filtered rotation matrix of the current frame are combined to obtain the optimized 3D pose data of the current frame. An optimized 3D bounding box is generated and output based on the optimized 3D pose data.

[0005] Optionally, the calculation of the filtered translation vector of the current frame specifically involves using an exponential smoothing filtering algorithm to perform a weighted calculation on the original translation vector of the current frame and the filtered translation vector of the previous frame stored in the translation buffer.

[0006] Optionally, the calculation formula for the exponential smoothing filter algorithm is: ; in, This is the filtered translation vector for the current frame. The smoothing coefficient controls the weighting of historical frame pose data. The translation vector of the previous frame, stored in the translation buffer. This is the original translation vector for the current frame.

[0007] Optionally, the smoothing coefficient The value range of is between 0 and 1, and the smoothing coefficient The value is adaptively adjusted according to real-time or smoothness requirements.

[0008] Optionally, the current frame's filtered rotation quaternion is calculated based on the current frame's rotation quaternion and the previous frame's filtered rotation quaternion stored in the rotation buffer. Specifically, a spherical linear interpolation algorithm is used to perform weighted fusion of the current frame's rotation quaternion and the previous frame's filtered rotation quaternion.

[0009] Optionally, the calculation formula for the spherical linear interpolation algorithm is: ; in, The filtered quaternion for the current frame. It is a spherical linear interpolation function. This is the rotated quaternion after filtering the previous frame. The rotation quaternion for the current frame. These are the preset smoothing weighting coefficients.

[0010] Optionally, the translation buffer and the rotation buffer are initialized empty buffers; for the first frame of the object to be optimized, the original translation vector of the first frame is stored in the translation buffer as the initial current frame filtered translation vector, and the rotation quaternion converted from the original rotation matrix of the first frame is stored in the rotation buffer as the initial current frame filtered rotation quaternion.

[0011] Optionally, the 2D to 3D boosting module is a component of a lightweight 3D pose estimation algorithm, and the input interfaces of the translation filtering branch and the rotation filtering branch are connected to the output interface of the 2D to 3D boosting module.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs a dual-branch independent filtering architecture for translation and rotation. It employs exponential smoothing filtering for the translation vector and spherical linear interpolation filtering based on quaternion transformation for the rotation matrix, achieving differentiated processing of two types of mathematically characteristic pose data. While ensuring the physical rationality of rotational orthogonality, it simultaneously optimizes position smoothness and pose continuity. By introducing a temporal caching mechanism, it establishes inter-frame pose correlations using the filtering results of the previous frame, effectively suppressing single-frame random errors caused by sensor noise and environmental interference, and significantly reducing displacement jitter and rotation jumps of the 3D bounding box in the temporal sequence. Furthermore, this invention's method is directly compatible with the output interface of existing lightweight 3D pose detection workflows, requiring no modification to the core structure of the original detection algorithm. While maintaining high inference speed, it provides more stable and accurate 3D pose estimation results for real-time applications such as autonomous driving and AR / VR. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 like Figure 1 As shown, this embodiment provides a 3D pose optimization method based on translation and rotation dual-branch temporal filtering, including: The 2D to 3D enhancement module is invoked to obtain the original 3D pose data of the object to be optimized in the current frame. The original 3D pose data includes translation vectors and rotation matrices. The translation vector is input into the translation filtering branch, and the filtered translation vector of the current frame is calculated based on the original translation vector of the current frame and the filtered translation vector of the previous frame stored in the translation buffer. The rotation matrix is ​​input into a rotation filtering branch that runs parallel to the translation filtering branch. The rotation matrix is ​​converted into a rotation quaternion for the current frame. Based on the rotation quaternion for the current frame and the previous frame's filtered rotation quaternion stored in the rotation buffer, the current frame's filtered rotation quaternion is calculated, and the current frame's filtered rotation quaternion is converted into a filtered rotation matrix. Update the translation buffer and the rotation buffer with the current frame filtered translation vector and the current frame filtered rotation quaternion, respectively. The filtered translation vector and the filtered rotation matrix of the current frame are combined to obtain the optimized 3D pose data of the current frame. An optimized 3D bounding box is generated and output based on the optimized 3D pose data.

[0019] The 2D to 3D enhancement module is invoked to obtain the original 3D pose data of the object to be optimized. The original 3D pose data includes two parts: translation vector T and rotation matrix R. A dual-branch temporal filtering architecture for translation and rotation is constructed. The architecture includes a translation filtering branch and a rotation filtering branch. The two branches are processed independently and in parallel, respectively performing temporal smoothing filtering on the translation vector T and the rotation matrix R. The original translation vector of the current frame is input into the translation filtering branch, and the filtered translation vector is calculated using the exponential smoothing filtering algorithm. At the same time, the filtered translation vector of the previous frame is stored in the translation buffer for use in the current frame filtering calculation. The original rotation matrix of the current frame is input into the rotation filtering branch, and the matrix to quaternion, spherical linear interpolation (SLERP), and quaternion to matrix operations are executed in sequence to obtain the filtered rotation matrix. At the same time, the rotation quaternion of the previous frame after filtering is stored in the rotation buffer for use in the current frame filtering calculation. The filtered translation vector and rotation quaternion are updated to the translation buffer and rotation buffer respectively, for use in the filtering calculation of the next frame; The filtered translation vector and the filtered rotation matrix are merged to obtain the complete filtered 3D pose data. Based on the filtered 3D pose data, the 3D bounding box of the object is reconstructed, and the optimized 3D bounding box with smooth temporal sequence is output, thus completing the 3D pose optimization.

[0020] Furthermore, the object to be optimized is an everyday rigid object supported by the lightweight 3D pose estimation algorithm, including but not limited to any one of shoes, chairs, and cups.

[0021] Furthermore, the exponential smoothing filter algorithm formula used in the translation filter branch is as follows: ,in, The smoothing coefficient controls the weighting of historical frame pose data. The translation vector of the previous frame, stored in the translation buffer. This is the original translation vector for the current frame.

[0022] Furthermore, the smoothing coefficient The range of values ​​is It can adaptively adjust according to the real-time and smoothness requirements of actual application scenarios. The larger the value, the higher the weight of the historical frame pose data, and the stronger the filtering and smoothing effect. The smaller the value, the higher the weight of the current frame's pose data, and the stronger the algorithm's real-time performance.

[0023] Furthermore, the specific operations of the rotating filter branch include: (1) Matrix to Quaternion: Convert the original 3×3 rotation matrix into a quaternion representation to obtain the rotation quaternion of the current frame. ; (2) Spherical Linear Interpolation (SLERP): Calls the spherical linear interpolation algorithm, combined with the filtered quaternion of the previous frame stored in the rotation buffer. Calculate the filtered rotation quaternion ,in The smoothing coefficient is the same as that of the translation filter branch; (3) Quaternion to Matrix: Convert the filtered quaternion to a matrix. Convert back to a 3×3 rotation matrix to obtain the filtered rotation matrix. .

[0024] Furthermore, the conversion between the matrix and the quaternion follows the mathematical conversion rules of rigid body rotation in three-dimensional space, ensuring the uniqueness and accuracy of the rotational posture representation.

[0025] Furthermore, the input interface of the dual-branch temporal filtering architecture is seamlessly connected to the output interface of the 2D to 3D enhancement module of the lightweight 3D pose detection algorithm. The output interface directly outputs the optimized 3D bounding box data, which can maintain good compatibility with the existing 3D pose detection process without requiring any modification to the core module of the original 3D pose detection algorithm, ensuring that the integrity and operating efficiency of the original detection process are not affected.

[0026] Furthermore, the translation buffer and rotation buffer are initialized empty buffers. For the first frame of data to be optimized, the original translation vector of the current frame is used. The translation vector after initial filtering is directly stored in the translation buffer, and the quaternion after the original rotation matrix of the current frame is transformed. The quaternion is directly stored in the rotation buffer after initial filtering.

[0027] Compared with the prior art, this embodiment has the following advantages and technical effects: (1) Independent filtering of translation and rotation branches to adapt to different data characteristics: This invention designs differentiated filtering strategies for the linear characteristics of translation vectors and the nonlinear rigid body rotation characteristics of rotation matrices. The translation branch adopts exponential smoothing filtering with simple calculation to ensure smooth position while taking into account real-time performance. The rotation branch adopts a filtering method that combines matrix and quaternion conversion with spherical linear interpolation to avoid the destruction of orthogonality caused by direct filtering of rotation matrix, ensuring the physical rationality and temporal continuity of rotation posture, and solving the technical problem that a single filtering strategy cannot optimize translation and rotation at the same time.

[0028] (2) Temporal state management, making full use of inter-frame association: By storing the filtering results of the previous frame through translation buffer and rotation buffer, the attitude association information between frames is established, so that the filtering calculation of each frame is combined with historical attitude data, effectively filtering out the random error of a single frame, greatly reducing the temporal jitter of the 3D bounding box, and significantly improving the attitude stability.

[0029] (3) Lightweight and efficient, compatible with the original algorithm architecture: The dual-branch filtering architecture of this invention has low computational complexity and can be directly and seamlessly connected with the general lightweight 3D pose detection process without modifying the core module of the original detection algorithm. It achieves 3D pose optimization while ensuring the high inference speed of the original detection process, ensuring efficient operation of real-time application scenarios such as mobile terminals and embedded terminals.

[0030] (4) Robust and adaptable to multiple practical application scenarios: The optimized 3D pose data effectively filters out false detection boxes caused by sensor errors and environmental interference. In autonomous driving and robot navigation, it can improve the accuracy of obstacle tracking, collision warning, path planning and high-precision map construction. In AR / VR, it can achieve precise fitting of virtual objects with real scenes, improve the interactive experience, and has a wide range of practical application value.

[0031] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0032] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0033] Example 2 This embodiment provides a 3D pose optimization method based on translation and rotation dual-branch temporal filtering, including: Step 1: Obtain the original 3D pose data of the object to be optimized, and complete the data source preparation before filtering; Step 1-1: Run a lightweight 3D object detection and pose estimation algorithm. Obtain the 2D clipping region of the object to be optimized through the object detection module and input the region into the pose conversion module. Step 1-2: The 2D to 3D enhancement module outputs the original 3D pose data of the object to be optimized. The original 3D pose data includes a translation vector T representing the position and a 3×3 rotation matrix R representing the orientation. Steps 1-3: Verify the validity of the original 3D pose data, remove invalid frame data with no detection results, and ensure the validity of subsequent filtering calculations; Step 2: Build a translation and rotation dual-branch time-series filter architecture, initialize the cache module and filter parameters, and complete the architecture and parameter preparation before filtering; Step 2-1: Build independent translation filtering and rotation filtering branches. The two branches are processed in parallel. The translation filtering branch only receives and processes the translation vector T, and the rotation filtering branch only receives and processes the rotation matrix R. Step 2-2: Initialize the translation buffer and rotation buffer as blank buffers. The translation buffer is used to store the translation vector after filtering in the previous frame, and the rotation buffer is used to store the rotation quaternion after filtering in the previous frame. Steps 2-3: Set the smoothing coefficient In this embodiment, the initial value is set to 0.7. This coefficient is a common parameter for both the translation filter branch and the rotation filter branch, and can be adaptively adjusted according to the real-time performance and smoothness requirements of the actual application scenario. Steps 2-4: Define the core calculation functions for exponential smoothing filtering, matrix-quaternion conversion, and spherical linear interpolation (SLERP) to ensure the accuracy of mathematical calculations in the filtering process; Step 3: Initialize filtering and cache storage for the first frame of data to be optimized, and complete the initial value setting of the filter; Step 3-1: Extract the original translation vector of the first frame and the original rotation matrix The input is fed into a two-branch time-series filter architecture; Step 3-2: The translation filter branch transforms the original translation vector Directly used as the translation vector after filtering in the first frame Stored in the translation buffer as historical reference data for translation filtering in the next frame; Step 3-3: The rotation filtering branch first processes the original rotation matrix. Convert to rotation quaternion It is directly used as the rotation quaternion after filtering the first frame. Stored in the rotation buffer as historical reference data for rotation filtering in the next frame; Steps 3-4: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error. Convert back to 3×3 rotation matrix ,merge and Reconstruct the optimized 3D bounding box of the first frame; Step 4: Perform dual-branch temporal filtering calculation on the data to be optimized in the nth frame (n≥2) to complete the single-frame attitude optimization and cache update; Step 4-1: Extract the original translation vector of the nth frame. and the original rotation matrix The inputs are respectively sent to the translation filter branch and the rotation filter branch; Step 4-2: The translation filtering branch calls the exponential smoothing filter function, combining it with the filtered translation vector of the previous frame stored in the translation buffer. According to the formula Calculate and obtain the translation vector after filtering in the nth frame. ; Step 4-3: The rotation filtering branch sequentially performs matrix to quaternion conversion and spherical linear interpolation. The quaternion-to-matrix operation yields the rotation matrix after filtering the nth frame. ; Step 4-4: Filter the translation vector Update to the translation buffer, and add the filtered rotation quaternion. Update to the rotation buffer, overwrite the original historical data, and provide a new historical reference for the next frame's filtering calculation; Steps 4-5: Merge and Reconstruct the optimized 3D bounding box of the nth frame to complete the 3D pose optimization of the nth frame; Step 5: Repeat the operation in Step 4 to perform dual-branch temporal filtering calculations on all frames in the video sequence to obtain the optimized 3D pose data and 3D bounding boxes for all frames. Step 6: Stitch the optimized 3D bounding boxes of all frames in time sequence to output a time-smooth 3D pose estimation result, thus completing the entire 3D pose optimization process.

[0034] In this embodiment, the detailed implementation steps are as follows: S1. In this embodiment, a chair is selected as the object to be optimized. A video sequence data package of the chair category is downloaded from the Google Objectron 3D object detection dataset. The package includes the original video, 3D annotation file and AR pose metadata. This data source can be directly replaced with video sequence data of various everyday rigid objects without modifying the filtering architecture and core calculation logic.

[0035] S2. Configure a Python virtual environment, resolve version compatibility issues of key libraries such as TensorFlow, OpenCV, and Protobuf, load a lightweight 3D pose detection model and run inference to obtain the original 3D pose data of each frame in the target object video sequence, namely the translation vector T (x, y, z 3D coordinates) and the 3×3 rotation matrix R. At the same time, write visualization code to realize the drawing of the original 3D bounding box and verify the validity of the original data.

[0036] S3. Construct a dual-branch temporal filtering architecture. The core of the translation filtering branch is the exponential smoothing filtering algorithm. The core logic of this algorithm is linear interpolation, which balances the weights of historical frame data and current frame data through a smoothing coefficient α. The formula is... In this context, 0.7 represents the weight of the filtering results of historical frames, and 0.3 represents the weight of the original data of the current frame. This ensures the smoothness of the translation position while avoiding the attitude lag problem caused by over-filtering.

[0037] S4. The rotation filtering branch addresses the nonlinear rigid body rotation characteristics of the rotation matrix by employing a combination of matrix and quaternion conversions and spherical linear interpolation. This avoids the orthogonality loss caused by directly filtering the rotation matrix. First, the original 3×3 rotation matrix... Convert to quaternion Quaternions, with their four components x, y, z, and w, represent rotational attitudes, effectively solving the singularity problem of rotation matrices. Then, the SLERP spherical linear interpolation algorithm is called, combined with the rotation buffer... According to the formula Calculate the filtered quaternion; finally, Convert back to 3×3 rotation matrix The entire process follows the mathematical transformation rules of rigid body rotation in three-dimensional space, ensuring the uniqueness and accuracy of the rotational posture representation.

[0038] S5. In this embodiment, the input interface of the translation and rotation dual-branch temporal filtering architecture is seamlessly connected to the output interface of the 2D to 3D enhancement module, directly receiving the translation vector and rotation matrix output by the module. The output interface directly outputs the optimized 3D bounding box data without modifying the core module of the original 3D pose estimation algorithm, maintaining high inference speed and running efficiency, and adapting to the real-time application requirements of mobile and embedded terminals.

[0039] S6. Quantitative analysis and visualization verification of 3D pose data before and after filtering are performed. The x-coordinate, y-coordinate, depth value, and displacement jitter value of the translation vector in each frame before and after filtering are extracted, and comparison curves are plotted. The results show that the translation vector coordinates and depth values ​​after filtering do not have obvious abrupt changes in the time frame, the curves are smooth and continuous, the average displacement jitter value decreases, and the filtering effect is significant. At the same time, the 2D projections of the 9 key points (1 center point plus 8 vertices) of the 3D bounding box before and after filtering are plotted on the video frame image using the graphics library. The visualization results show that the optimized 3D bounding box does not have obvious jumps and always accurately fits the chair target in the continuous video frames, proving that the method of the present invention effectively solves the temporal jitter problem of the existing 3D pose estimation method.

[0040] S7. Smoothing coefficient in this embodiment It can be flexibly adjusted according to the actual application scenario: if applied to AR / VR virtual-real fusion scenarios, where high smoothness is required, it can be... Adjusting the value to 0.8-0.9 increases the weight of historical frame data and enhances the smoothing effect; if applied to real-time collision warning scenarios in autonomous driving, where real-time performance is critical, the value can be adjusted accordingly. Adjust the value to 0.5-0.6 to increase the weight of the current frame data and ensure the real-time performance of attitude estimation.

[0041] In this embodiment, the optimization model selects the exponential smoothing filter algorithm as the translation filter branch and the matrix-quaternion interconversion combined with spherical linear interpolation algorithm as the rotation filter branch. The verification module uses Euclidean distance to calculate the error, and the output is the optimized translation vector and rotation matrix corresponding to the pose data of that frame. The optimization effect is verified by the inter-frame jitter value. Here, the pose data of that frame refers to the target temporal frame data currently participating in the training, because the original labeled data has standard values ​​and does not require optimization.

[0042] This embodiment employs a differentiated filtering strategy to process translation vectors and rotation matrices separately within a dual-branch temporal filtering architecture, adapting to their different mathematical characteristics. By minimizing filtering errors, it achieves synchronous optimization of translation position and rotation attitude. Secondly, a temporal state management mechanism is added to the cache module, utilizing inter-frame attitude correlation information to filter out single-frame random errors, ensuring the temporal continuity of 3D attitude estimation. Finally, through seamless integration with a general lightweight 3D attitude detection workflow, it maintains the high inference speed of the original algorithm while achieving attitude optimization, ensuring the method's engineering practicality.

[0043] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A 3D pose optimization method based on translation and rotation dual-branch temporal filtering, characterized in that, include: The 2D to 3D enhancement module is invoked to obtain the original 3D pose data of the object to be optimized in the current frame. The original 3D pose data includes translation vectors and rotation matrices. The translation vector is input into the translation filtering branch, and the filtered translation vector of the current frame is calculated based on the original translation vector of the current frame and the filtered translation vector of the previous frame stored in the translation buffer. The rotation matrix is ​​input into a rotation filtering branch that runs parallel to the translation filtering branch. The rotation matrix is ​​converted into a rotation quaternion for the current frame. Based on the rotation quaternion for the current frame and the previous frame's filtered rotation quaternion stored in the rotation buffer, the current frame's filtered rotation quaternion is calculated, and the current frame's filtered rotation quaternion is converted into a filtered rotation matrix. Update the translation buffer and the rotation buffer with the current frame filtered translation vector and the current frame filtered rotation quaternion, respectively. The filtered translation vector and the filtered rotation matrix of the current frame are combined to obtain the optimized 3D pose data of the current frame. An optimized 3D bounding box is generated and output based on the optimized 3D pose data.

2. The method according to claim 1, characterized in that, The calculation of the current frame's filtered translation vector specifically involves using an exponential smoothing filtering algorithm to perform a weighted calculation on the original translation vector of the current frame and the filtered translation vector of the previous frame stored in the translation buffer.

3. The method according to claim 2, characterized in that, The calculation formula for the exponential smoothing filter algorithm is as follows: ； in, This is the filtered translation vector for the current frame. The smoothing coefficient controls the weighting of historical frame pose data. The translation vector of the previous frame, stored in the translation buffer. This is the original translation vector for the current frame.

4. The method according to claim 3, characterized in that, The smoothness coefficient The value range of is between 0 and 1, and the smoothing coefficient The value is adaptively adjusted according to real-time or smoothness requirements.

5. The method according to claim 1, characterized in that, The current frame's rotated quaternion is calculated based on the current frame's rotated quaternion and the previous frame's filtered rotated quaternion stored in the rotation buffer. Specifically, a spherical linear interpolation algorithm is used to perform a weighted fusion of the current frame's rotated quaternion and the previous frame's filtered rotated quaternion.

6. The method according to claim 5, characterized in that, The calculation formula for the spherical linear interpolation algorithm is as follows: ； in, The filtered quaternion for the current frame. It is a spherical linear interpolation function. This is the rotated quaternion after filtering the previous frame. The rotation quaternion for the current frame. These are the preset smoothing weighting coefficients.

7. The method according to claim 1, characterized in that, The translation buffer and the rotation buffer are initialized empty buffers; For the first frame of the object to be optimized, the original translation vector of the first frame is stored in the translation buffer as the initial current frame filtered translation vector, and the rotation quaternion converted from the original rotation matrix of the first frame is stored in the rotation buffer as the initial current frame filtered rotation quaternion.

8. The method according to claim 1, characterized in that, The 2D to 3D boosting module is a component of the lightweight 3D pose estimation algorithm, and the input interfaces of the translation filtering branch and the rotation filtering branch are connected to the output interface of the 2D to 3D boosting module.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images